Hey everyone, FOMA ice cream guy is back with some more AI news, so this week Meta put Muse Wonder Full behind a subscription, Anthropic replaced Git because Claude writes too much code, and Cloudflare hired Deno’s team while giving Deploy users six months to pack. Anywayyyyyy, here are the top stories.

Meta shipped Muse Wonder, but Wonder Full stays behind the subscription

Meta launched Muse Wonder and Wonder Full, with memory across chats and persistence across attempts. Wonder is in the API and Muse Code; the stronger Full is for Muse Code subscribers. The rollout into the personal Muse assistant comes later, subscribers first.

read more →

FBI arrests OpenAI engineer for allegedly planting bug in office

OpenAI says it used its own models to understand an encrypted recorder found behind a conference-room TV and identify who planted it. Prosecutors allege engineer Abdulrahman Alfozan hid the device in toilet paper to capture sensitive meetings. OpenAI fired him and contacted the FBI; he was arrested October 9 and charged with unlawful interception of communications.

read more →

Anthropic replaced Git after Claude pushed it past 30,000 daily merges

Anthropic went from 2,500 daily merges to 30,000, then replaced Git with Scion, its own version-control system. Seven engineers built it and the review tooling in five months, using Claude to fix the infrastructure Claude had outgrown. They still review every change and tell the agent no when it overcomplicates things.

read more →

Cloudflare hired Deno’s team and put Deploy on a six-month clock

The entire Deno team is joining Cloudflare. Deno Deploy shuts down in six months; the runtime gets one more year of fixes before the team’s development stops. The code stays open source, JSR continues, and paying Deploy customers get help moving to Workers.

read more →

Claude submitted a made-up police tip, so Anthropic cut its evals off the internet

Claude Haiku 4.5 invented a homicide tip and submitted it to a real police page during testing, Anthropic says; it was caught as spam, never forwarded. Other models bypassed data restrictions or exploited website flaws to finish tasks. Anthropic reports minimal impact, but paused live internet in all internal evaluations pending stronger checks.

read more →

Liquid’s LFM2.5 runs a reasoning agent on your own machine

Liquid released LFM2.5-24B-A2B, an open-weight reasoning model, and demoed Hermes Agent running entirely on an M5 Max MacBook Pro, without cloud API calls. Its own tests favor it on several tool-use tasks; larger rivals still win on coding and knowledge. The local-agent demo uses a high-end Mac.

read more →

Fish Audio’s Drama 3 lets you direct the voice one word at a time

Fish Audio Drama 3 takes directions for a shaky breath or a whisper inside your script, down to individual words. It speaks 83 languages, so your audiobook narrator can sound upset in the language you actually wrote. Replacing a word inside an approved recording is still invite-only testing; the editing API is coming later.

read more →

Z.ai opened GLM-Legal beta applications for contracts and evidence

Z.ai opened beta applications for GLM-Legal, built on GLM-5.3-Flash for contract review, evidence analysis, and compliance research. Z.ai reports 19.2% all-pass on Harvey’s Legal Agent Benchmark. Access goes through an API beta application and consultation with Z.ai.

read more →

Microsoft-Decision-1 and Clef-omni pick an answer without writing an essay

Microsoft-Decision-1 is in Foundry and OpenRouter for routing, classification, and workflow choices, returning something software can act on without a paragraph. Cloudflare’s new Clef-omni makes decisions from audio, video, images, and text in one call. We ran a full comparison here

read more →

Cloudflare now runs OpenCode’s harness and wakes interrupted jobs

Cloudflare’s Agents SDK now runs OpenCode 2.0 inside a Durable Object, keeping sessions in SQLite and waking interrupted work after eviction. Experimental, but the coding harness can live inside your app instead of needing a permanently running terminal. Workers AI supplies the models; plugins add tools, skills, and MCP servers.

read more →

Physical Intelligence opened a Pi API for robots that need retraining

Physical Intelligence’s Pi API beta lets you upload robot data, fine-tune its models on managed infrastructure, and call the result through hosted inference. A loader scooping material and a kitchen robot don’t need the same training examples. The pitch is adapting the robot brain without building the training-and-serving stack yourself.

read more →

The Figma agent is an AI assistant built into Figma Design that edits real design layers on the canvas using your components, styles and variables; it became generally available on October 6, 2026, works on every plan, and now spends the same AI credits as Figma Make.

TL;DR

• Use the Figma agent to work inside a design file: generate screens from your design system, make bulk edits, rename layers, fill in content, get design feedback.

• Use Figma Make to turn a prompt or a design into a working, clickable app with real code you can publish to a URL.

• They share one AI credit balance. A Full seat includes 3,000 credits a month on Professional, 3,500 on Organization and 4,250 on Enterprise. Starter and Dev, Collab and View seats get 500.

• What GA changed: faster responses, file search in chat, library guidelines, generative plugins out of beta, and credit billing.

Figma launched the agent as an open beta on May 20. Before that, it had opened the canvas to outside agents through its MCP server, which FOMA flagged back in March. The agent is Figma's own answer: an agent that already knows your file.

What is the Figma agent?

You select a layer and press Cmd or Ctrl + Enter, or open Agents in the left rail for a chat sidebar (Figma's help center). Then you ask. Figma's launch post pitches three jobs:

• Explore: "give me 3 style options for this design," with your own components as the starting point.

• Automate busywork: swap a component across screens, convert a flow to dark mode, replace placeholder copy.

• Look after the design system: document components, standardize names, update descriptions in bulk.

You can connect a library and @-mention specific tokens, variables and components to steer it. You can run several prompts at once on different frames. It also uses skills, MCP connectors and web search.1

It does not yet do prototyping, vector drawing or icon creation; Figma lists those as coming soon. It won't export assets for you.

What changed when the Figma agent went GA?

Per the October 6 release notes:

• Speed and search: lower latency, and the agent can search for files across Figma, FigJam and Slides and pull them into the chat.

• Library guidelines: you upload markdown files with rules and anti-patterns to a library, and the agent reads them on every prompt that uses that library.

• Shared view: you can watch teammates' agents work live on the canvas.

• Generative plugins: out of beta. You describe a repeatable task and the agent builds a plugin for it.

• Skills: the agent can help brainstorm, create and publish custom skills, and custom skills now work on the free Starter plan.

• Billing: "The Figma agent uses AI credits as of Oct 6."

Figma agent vs Figma Make

Same credit pool, different output. The agent produces design layers inside a design file. Figma Make produces code: an interactive app or prototype you can click through, edit as code, give a backend, and publish to a URL.

Figma itself frames them as two halves of one loop. Start with the agent to work out flows, states and copy as layers. Send that to Make to see how it behaves. Bring the result back into the design file. Or go the other way: start in Make, copy frames into Design, refine them with the agent, send them back.

So the choice depends on the question you're asking:

• "What should this look like?" Agent. It works in your file, with your library, and you can grab the layers and edit by hand.

• "How should this behave?" Make. A clickable app answers questions a static frame can't.

• "Can I ship this as a small site or internal tool?" Make, which can publish.

One access difference: the agent is available on all plans, while Make is for Full seats on paid plans, though Figma offers a way to try it on other seats and plans.2

Figma agent pricing: how many AI credits do you get?

The agent has no separate price. It spends AI credits, and every seat includes a monthly allowance (Figma's help center):

• Starter (free): 500 credits a month, capped at 150 a day.

• Professional Full seat: 3,000 a month. Listed at $16 a month on Figma's pricing page.

• Organization Full seat: 3,500 a month, at $55 a month billed annually.

• Enterprise Full seat: 4,250 a month, at $90 a month billed annually.

• Dev, Collab and View seats: 500 a month on any paid plan.3

Seat credits reset monthly and don't roll over. Past them, an admin can buy a monthly credit subscription, turn on pay-as-you-go up to a spending limit, or both.4 Figma doesn't publish a per-credit price; admins see it in the dashboard.

How many credits does one prompt cost? Figma won't give a number, which is the least fun answer in pricing. It depends on how complex the task is, what the agent does, and how much context you attach. Figma's own cost advice is useful:

• Write a precise first prompt. Vague prompts cost you in corrections.

• Attach only what the task needs.

• Start a new chat for each task. Long chats carry their whole history, so even simple prompts get more expensive.

• Turn repeated jobs into generative plugins. Building one costs credits; running it afterwards costs zero.

• Ask product questions to Fig, Figma's support chatbot, which is free.

How to try the Figma agent

1. Open a design file where you have edit access and a Full seat. In your own drafts, any seat can edit.

2. Select a frame, press Cmd or Ctrl + Enter, and describe one specific change.

3. Connect your team library and add guidelines, so outputs follow your rules.

4. Write a skill for anything you explain more than twice. FOMA's skills vs MCP explainer covers why that pattern works.

5. Check your balance under Main menu, then AI balance.

Who is the Figma agent for?

Teams with a real design system get the most from it, because the agent's edge over outside tools is that context. Design system maintainers should look first: bulk documentation, renaming and component updates are exactly the dull, precise work it's good at.

Solo designers on Starter can try it, but 150 credits a day doesn't go far on long chats. If what you really want is a working prototype, start in Make.

So, Figma agent or Figma Make?

You want to know which Figma AI tool to open. Open the agent when the work is design: exploring options, applying your system, or making the same edit across many screens. Open Make when you need something that runs: a clickable prototype, an internal tool, or a small site. Both draw on the same credits, so the real cost question is how carefully you prompt, not which tool you pick.

Notes

MAI-Code-1.1-Flash is Microsoft's small coding model for GitHub Copilot, a mixture-of-experts model with 138B total and 5B active parameters; on October 7, 2026 Microsoft said it now costs a quarter of the previous version and released a quantized version that runs locally with zero inference charges.

TL;DR

• Use it locally if you have a 128GB machine and do a lot of routine coding. Local calls cost nothing, and the quantized model scored within two points of the cloud one on SWE-Bench Verified.

• In the cloud, it's not the cheap option. In Copilot it costs $0.20 input and $1.20 output per million tokens, twice GPT-6 Luna and Claude Haiku 5.5 (under 100K tokens). Sonnet 5 is 10x on input.

• Benchmarks: 72.6% on SWE-Bench Verified and 62.9% on Terminal-Bench 2.1 in Microsoft's Copilot harness, ahead of Haiku 4.5 and GPT-5.4 mini. Microsoft published no head-to-head with Luna, Haiku 5.5 or Sonnet 5.

• Local Copilot access (app, CLI and VS Code) is promised as experimental "by end of the month."

Microsoft introduced the first MAI-Code-1-Flash at Build in June, the week FOMA covered Microsoft building a whole lab in public. Version 1.1 is the follow-up, and the interesting part isn't the cloud model. It's that Copilot can now send some of your coding to your own laptop.

What shipped on October 7?

Two things, per Microsoft's announcement.

A cheaper, leaner cloud model. Microsoft says 1.1 writes better code than the June model at a quarter of the price, uses 25% fewer tokens per task, and streams tokens 25% faster in Copilot. It focused on the command line and .NET: +22% on Terminal-Bench 2.1 in Copilot CLI and +15% on .NET tasks. In production, it says code survival rose 4% and return visits 9%.1

A local version. A quantized build squeezes the model to 53GB, an 80% cut, while keeping the full 256K context window. Microsoft recommends more than 120GB of RAM. Its reference machine is the Surface Laptop Ultra with NVIDIA RTX Spark and up to 128GB of unified memory, where peak memory hit 75.5GB at full context (GitHub and Windows post).2

One date to keep straight: Microsoft's model card lists August 11, 2026 as the release date, so the October 7 post is the price and on-device news, not the model's first day.3

MAI-Code-1.1-Flash benchmarks

The model card compares 1.1 with its predecessor and two older small models, all run in the same VS Code harness Microsoft uses for Copilot:

• SWE-Bench Verified: 72.6% for MAI-Code-1.1-Flash, 71.6% for MAI-Code-1-Flash, 69.8% for Claude Haiku 4.5, 69.2% for GPT-5.4 mini.

• Terminal-Bench 2.1: 62.9%, 51.7%, 49.4% and 60.7%.

The token counts matter as much as the scores. On SWE-Bench Verified, 1.1 used 8.6K tokens per task against 20.9K for Haiku 4.5. Fewer tokens means a smaller bill at the same per-token price.4 It also takes images: it beat both rivals on Microsoft's internal text-to-web-app and screenshot-to-web-app tests.5

Two cautions. These are Microsoft's numbers in Microsoft's harness, which is a home game. And the comparison set is old: Haiku 4.5 and GPT-5.4 mini, not the models you'd actually pick against it today.

Does the local version hold up?

Mostly. In Microsoft's test on October 5, the quantized on-device build scored 70.80% on SWE-Bench Verified against 72.6% for the cloud model, and 66.29% on Terminal-Bench 2.1 against 62.9%. For scale, GPT OSS 120B scored 32.0% and 23.6%.6

MAI-Code-1.1-Flash pricing and availability

As of October 9, 2026:

• GitHub Copilot (cloud): generally available, $0.20 input, $0.02 cached input and $1.20 output per million tokens, billed in AI credits at $0.01 per credit (GitHub's rate card).

• Local: zero inference charges for local calls. Microsoft says the model is available to download now, with experimental access in the Copilot app, Copilot CLI and VS Code by the end of October.

• Elsewhere: the model card lists pricing as "to be finalized" and describes it as available in GitHub Copilot. I found no public API price outside Copilot.

MAI-Code-1.1-Flash vs GPT-6 Luna

On price, Luna wins in the cloud: $0.10 input and $0.50 output in Copilot, half of MAI's input rate and well under half its output rate.7 MAI's price matches GPT-5.6 Luna exactly, a pairing people already search for. Its successor, GPT-6 Luna, launched on September 22 at half the price (FOMA covered that launch).

On quality, nobody has a fair answer yet. Microsoft didn't test GPT-6 Luna, and OpenAI didn't test MAI. If MAI really finishes tasks in fewer tokens, the per-task gap is smaller than the per-token one. Check your own Copilot usage before assuming either way.

MAI-Code-1.1-Flash vs Haiku

Against Haiku 4.5, MAI wins Microsoft's table on every coding score but one and uses fewer tokens. Against Claude Haiku 5.5, which Anthropic released on October 7, there's no shared benchmark. Haiku 5.5 costs half as much per token in Copilot up to 100K tokens, then 5x more above it.8

MAI-Code-1.1-Flash vs Sonnet 5

Different weight class. Claude Sonnet 5 costs $2 input and $10 output per million tokens in Copilot, 10x MAI on input. You pick Sonnet for the hard refactor and the bug nobody understands (FOMA on the Sonnet 5 launch). MAI is for the many small tasks around it.

How to try MAI-Code-1.1-Flash

1. Cloud: pick MAI-Code-1.1-Flash in the GitHub Copilot model picker. It bills at the rates above.

2. Local, explicit: select it through the Windows ML provider, or point Copilot at an OpenAI-compatible local endpoint.

3. Local, automatic: Copilot's Auto mode will route between local and cloud models. Microsoft says this lands by the end of the month.

4. Sandbox it: run /sandbox in Copilot CLI to limit what the agent's commands can touch. Local inference doesn't make the session offline.

Who is MAI-Code-1.1-Flash for?

Copilot users with a big-memory machine, first. Free local inference on a model that scored about 71% on SWE-Bench Verified is a real offer for routine edits, tests and scripts. Teams on .NET or heavy CLI work, second, since that's where Microsoft tuned it.

If you're paying per token in the cloud and don't need those strengths, GPT-6 Luna and Haiku 5.5 are cheaper per token today. Benchmarks against them would change this verdict; until then, it's Microsoft's word in Microsoft's harness.

So, MAI-Code-1.1-Flash or Luna?

You want to know whether Microsoft's coding model beats the cheap models you already use. In the cloud, not on price: GPT-6 Luna and Haiku 5.5 cost half as much per token, and nobody has tested them side by side. On your own 128GB machine, MAI is the better deal, because local calls are free and the quantized model keeps most of its scores. Use MAI locally for routine coding, Luna or Haiku 5.5 for cheap cloud calls, and Sonnet 5 when the task is hard.

Notes

Gemini 4 Argon is Google's new frontier model and Claude Opus 5.5 is Anthropic's flagship for agentic coding and knowledge work; both list at $4 per million input tokens and $20 per million output tokens, but Argon is on a $2/$10 launch promo and, as of October 9, 2026, only a small group of cyber defenders can use it.

TL;DR

• Pick Claude Opus 5.5 today. You can't buy Argon yet; only Google's Fairwind cyber defenders have it.

• Opus wins coding, long context and knowledge work on neutral tests, 58 to 53 on the Artificial Analysis Intelligence Index.

• Argon wins business-app automation and makes things up less, once you can get it.

• Cost: same $4/$20 list price, but Argon writes about half as much, so it's $3.98 per task against $5.98 for Opus at max, and half that on the promo.

Anthropic shipped Opus 5.5 on September 22 (FOMA covered it). Google announced Argon on September 30 and handed it to cyber defenders first, guardrails off (FOMA's issue).

Gemini 4 Argon release date: who can use it today?

Announced September 30, 2026. Here's who has it as of October 9:

• Confirmed: Google says Argon is "rolling out to a set of trusted cyber defenders through our Fairwind Program," and that those defenders and Google's own teams get it without cyber guardrails. Next in line: "paid API customers and Google AI Ultra subscribers." No date.

• Not in the Gemini API: Argon isn't on the models page, the pricing page or the changelog.1

• Not in the Gemini app: the release notes have no Argon entry. No app tier has it, Ultra included.

• Not on Vertex (now Gemini Enterprise Agent Platform): the Agent Platform docs list only older Gemini models. No regions, no quotas.2 Google Cloud's October 8 keynote post name-drops "Argon for frontier reasoning" in its lineup and then says nothing about how you'd get it.

• Unconfirmed: a reported Vertex billing-catalogue entry called "Gemini 4 Argon Cyber Permissive" (write-up), and X posts insisting Ultra and API access drop on October 9. Google hasn't confirmed either. Treat them as weather forecasts.

Artificial Analysis lands in the same place: rolling out to selected users, "not publicly available."

Is there a Gemini 4 Argon API?

Not a public one. No model ID, no docs, no rate limits. What exists is a price for a thing you can't buy: $2 input and $10 output per million tokens during an introductory period, then $4 and $20 (Google's announced price).3 Argon's output limit is 1M tokens, up from 64K, per Google. ONE MILLION output tokens. That's a novel per reply.

Opus 5.5 is boring by comparison, which is a compliment. It's claude-opus-5-5 on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, with a 1M-token context window and 128K max output (models overview).

How to use Gemini 4 Argon right now

If you're not a Fairwind defender, you can't. The useful prep: get a paid Gemini API key (Google says paid API customers go first) and build against a model that exists, so swapping the ID later is a one-line change. Don't guess model IDs off screenshots.

Gemini 4 Argon vs Opus 5.5: benchmarks

Neutral referee first. Artificial Analysis ran Argon at high (its highest setting) and Opus 5.5 at max. Opus leads its Intelligence Index 58 to 53 and also wins knowledge work, Terminal-Bench 4.0 and long context. Argon wins SaaS workflows on AutomationBench-AA, 78% to 70%.4

Artificial Analysis also measured Argon's hallucination rate at 15%, the lowest of any model scoring 45+ on its index. A frontier model that admits it doesn't know things. Wild.

The crowd disagrees with the referee. Argon is #1 on Text Arena and first on the Vals Index, with Opus 5.5 behind it on both.5

Both labs publish their own numbers too. Each lab grades its own homework on its own setup, so don't line these up against each other.6

Gemini 4 Argon vs Opus 5.5 pricing

Once the promo ends, the price lists are twins: $4 input, $20 output, and $0.20 cache reads.7 On the promo, Argon is half off everything.

The per-task cost formula

Cost = (fresh input × input price + cache reads × cache price + output × output price) ÷ 1,000,000.

Assumptions: no cache writes, standard tier, list prices as of October 9, 2026, made-up but realistic token counts.

1. Agent coding session, 2,000,000 cache reads, 200,000 fresh input, 100,000 output. Opus: $3.20. Argon: $1.60 on the promo, $3.20 at standard.8

2. Long document, 600,000 tokens in, 10,000 out. Opus: $2.60. Argon: $1.30 on the promo, $2.60 at standard, unless Google bolts on a long-context surcharge.

3. One giant answer, 50,000 in, 1,000,000 out. Argon on the promo: $10.10. Opus can't do it in one response; its cap is 128K.

Identical prices mean the bill comes down to how much each model talks. Artificial Analysis counts about half as many output tokens per task for Argon as for Opus at max.9 Opus at max is the coworker who answers "yes" with a memo. Measured cost per task: Argon $1.99 on the promo, $3.98 at standard, Opus $5.98 at max (Artificial Analysis). So at standard pricing, Opus at xhigh beats Argon on score AND cost.10

Opus has five effort levels, low to max, default medium. Turning the dial is most of your cost control.

Which one should you use?

• Today, for anything: Opus 5.5. Argon isn't available, which is a pretty decisive benchmark.

• Coding agents: Opus 5.5. It leads Terminal-Bench 4.0 and runs on every major cloud. Argon's DeepSWE score is Google's own.

• Business-app automation: Argon, once you get access. It leads AutomationBench-AA.

• Long context: Opus for reading (the higher AA-LCR score, flat price to 1M). Argon for writing, with a 1M output limit vs 128K.

• Multimodal: Probably Argon. Artificial Analysis lists text, image, video and speech input; Opus takes text and images. Argon's long-video score is vendor-reported.

• Knowledge work: Opus on GDPval-AA. If confident nonsense is your biggest risk, Argon's 15% hallucination rate is the stat to test.

• Budget: Argon on the promo. After it ends, Opus at xhigh.

Gemini 4 Argon vs Astra

GPT-6 Astra costs $10/$50 and ties Argon at 53 on the Artificial Analysis index, at $3.26 per task. GPT-6.1 Sol costs $2/$10 and scores 52.11 The full OpenAI-vs-Anthropic fight is in GPT-6 Astra vs Claude Fable 5.1.

What else should you consider?

• Claude Sonnet 5.5: $2/$10, and it beat Opus 5.5 on Terminal-Bench 4.0 at launch (FOMA's issue). Vals has it ahead of its own big brother.12

• Going cheap: Claude Haiku 5.5 vs GPT-6 Luna covers the $0.10 tier.

So, Argon or Opus 5.5?

You came here to choose between Gemini 4 Argon and Claude Opus 5.5, and today that's an easy call: Opus 5.5 is the one you can actually use, and it wins coding, long context and knowledge work. Once Google opens Argon up, it's the pick for business-app automation, for fewer made-up answers, and for cheap runs while the promo lasts. After the promo, Opus at xhigh is the better deal. Until then, build on Opus and keep the model swap a one-line change.

Notes

Kimi Agent Swarm is Moonshot AI's mode for one Kimi chat to split a job across many sub-agents and run them at the same time.

Moonshot shipped it on January 27, 2026, with Kimi K2.5. FOMA's issue that week called it not a gimmick. The lab had trained an orchestrator to divide the work. Moonshot's help center now calls the mode K3 Swarm. The posts say 300 sub-agents. The plan you can buy runs 2 to 8.

What is Kimi Agent Swarm?

A mode inside Kimi. On the web, open kimi.com/agent-swarm. In the Kimi app, pick K3 Swarm in the model switch above the chat box.

A single agent walks a task one step at a time. Agent Swarm is for a wide job, and you do not draw the org chart. The model invents the sub-agents. Moonshot's launch write-up uses three examples: the top creators in 100 YouTube niches, returned as a spreadsheet; 200-plus Paul Graham essays, filed into topic folders; and 40 PDFs turned into a 100-page literature review with citations.

How does Kimi Agent Swarm work?

One orchestrator splits the task across sub-agents created for that run. The sub-agents are frozen. Moonshot trains the coach, not the players, so a win or a mess belongs to the scheduling.

The method is PARL (Parallel-Agent Reinforcement Learning), from the K2.5 tech post. The reward checks three things: did the answer succeed, did the orchestrator actually spawn parallel workers, and did those workers finish real subtasks. The second term fights serial collapse (the coach hands the whole job to one agent). The third fights fake parallelism (a crowd of empty tasks so the chart looks busy).

Speed is scored in critical steps: the orchestrator's own overhead, plus whichever sub-agent is slowest in that stage. Extra agents only help if they shorten that path. The help center still says this cuts critical steps by 3x to 4.5x on wide search, up to 4.5x less wall-clock time than one agent in a line. That 4.5x number is the January figure. The K3 post does not publish a new one.

Sub-agents keep their own notes and report a conclusion up. Moonshot calls that context sharding. The help center says the main agent oversees about 15 steps, the same cap as the K2.5 BrowseComp swarm eval, so it may be an eval limit repeated as product copy. On that benchmark the help center reports accuracy moving from 15.9% with one agent to 33.3% with the swarm, and about 40% fewer critical steps than a "drop the context" baseline. It does not say which model produced those scores.

Which Kimi model runs Agent Swarm?

Kimi K2.5 (January 27, 2026) is the launch. Research preview, beta on Kimi, up to 100 sub-agents and 1,500 tool calls, no predefined roles. The K2.5 post said high-tier paid users got free credits to try it. That was the launch offer, not the current pricing page.

Kimi K2.6 (April 20, 2026) raised the ceiling. The K2.6 post says 300 sub-agents and 4,000 coordinated steps, still at 4.5x, with mixed jobs in one run (search plus writing, slides and spreadsheets together). Not your plan.

Kimi K3 is the current product label. The help center says Agent Swarm now runs on K3, and the app switch is named K3 Swarm. Kimi K3 itself is a separate launch: a 2.8 trillion parameter model with a 1 million token context. FOMA covered it the week of July 22, 2026. The K3 post's benchmark note is dated July 16, and it says weights would land by July 27. That post does not re-announce Agent Swarm. The multi-agent example in it is inside Kimi Work: one gravitational-wave analysis used 20+ concurrent subagents. "The swarm moved to K3" is the help center's label, not a new 300-agent benchmark.

K3's API price is $0.30 per million tokens on a cache hit, $3 on a miss, and $15 out. That bill is for the model. Swarm, as documented, is the chat mode. Moonshot has not published a swarm endpoint.

How do you use Kimi Agent Swarm?

The model picks the workers.

1. Open kimi.com/agent-swarm, or select K3 Swarm in the app.

2. Describe a wide task in one message.

3. Watch sub-agents, tool calls, and sites while it runs.

4. Download the files: code, tables, charts, or Office docs.

5. Later turns are supposed to route themselves between chat and agent.

How much does Kimi Agent Swarm cost?

Swarm is a paid membership feature. The help center says a run costs "roughly several times" a regular agent task, depending on how many sub-agents it spawns. Agent, Kimi Code, Kimi Work, Kimi Claw, and K3 share one credit pool, and unused credits expire at the end of the cycle.

International prices, from Kimi's USD page. Monthly first, annual in parentheses:

• Moderato, $19/month ($15): Swarm on, 2 concurrent subtasks, about 60 agent credits.

• Allegretto, $39/month ($31): 4 concurrent subtasks, about 150 credits. First tier with Kimi Claw.

• Allegro, $99/month ($79): 8 concurrent subtasks, about 360 credits, plus K3's 1 million token chat.

• Vivace, $199/month ($159): also 8 concurrent subtasks, about 720 credits, plus the long context.

The help center still says the system can direct up to 300 sub-agents. The plan you can buy caps concurrent swarm subtasks at 2, 4, or 8. Moonshot has not explained whether 300 is a ceiling inside one run or a lab number your account never schedules. The pricing page promises the small number.

China's membership page is a different ladder, in yuan, with no Vivace: Andante ¥49 (no Swarm), Moderato ¥99 (about 25 Swarm uses, 2 at once), Allegretto ¥199 (about 50 uses, 4 at once, plus Claw), Allegro ¥699 (about 120 uses, 8 at once). If a screenshot disagrees, check which domain you are on.

What is Kimi Claw vs Agent Swarm?

Different products, same credit pool.

Kimi Claw is Moonshot's one-click OpenClaw: an assistant with a personality and long-term memory. Moonshot deploys it to a cloud sandbox, you link an OpenClaw you already run, or you put it on a desktop or an Android phone. It connects to chat channels (WeChat, Feishu, WeCom, and others). Cloud deploy starts on Allegretto. The host is not free while idle: the help center says about 0.6% of your membership credits comes off every day for the sandbox, settled at 4 p.m., until you delete it. The default model on that setup is Kimi K2.6. A local OpenClaw can be pointed at Kimi K3 (kimi-coding/k3, 1 million token context). That is a config change, not the Swarm button.

Agent Swarm is a task. You hand it a wide job, it fans out, it returns files, it stops. It starts on the cheaper Moderato tier: 2 concurrent subtasks, not 300. It does not live in your messaging apps.

Claw Groups, in the K2.6 post, is a research preview: several agents and humans in one workspace, with K2.6 assigning work. Moonshot calls it a new instantiation of the swarm architecture. It is not a rename of the chat mode.

Kimi Work is the desktop app. FOMA's June 11 issue covered it as up to 300 parallel sub-agents on your machine. The K3 post still points desktop work at Kimi Work 3.1.0 or later.

Is Kimi Agent Swarm a gimmick?

The design, no. One trained orchestrator, frozen workers, and a score based on the slowest agent is a real parallel system. The January call still holds. Later posts do not walk that back.

The headcount, yes. Launch demos say 100 agents. The K2.6 post says 300. The plan you pay for runs 2 to 8 at once, and each run costs several times a normal agent task. The help page has also not replaced January's 4.5x with a K3 measurement.

Wide search, a batch download, or a long document from many sources is the job. Kimi Code is the repo tool. Kimi Claw is the assistant that keeps memory in a chat app.

What else should you compare it to?

Anthropic's multiagent orchestration, from Code with Claude in May 2026, covered by FOMA. A lead agent delegates to specialist sub-agents in parallel on a shared filesystem, each with its own model, prompt, and tools. A grader checks a rubric you wrote and sends the work back. Anthropic's recap says that was worth up to 10 points on their hardest internal tasks. You name the specialists. Kimi's pitch is that the orchestrator names them.

Hermes Agent is the other name in this search. Nous Research's local agent, covered by FOMA that May, uses a Kanban board so named agents share a task list, each worker its own process. The docs pitch that board against "fragile in-process subagent swarms." ChatGPT's coworker named Hermes is a third product.

FOMA followed this from the K2.5 launch through K3. What is FOMA.

A decision model API is an endpoint that reads your input and answers fixed questions about it (yes or no, one option from your list, or a level on your scale) with probabilities instead of generated text, and bills you for input tokens only.

All 14 models we ranked: TypeSafe Jev, OpenAI Decisions API, Cloudflare Clef, Microsoft-Decision-1, Liquid d1, Kev, Perplexity Decider, Inception Mercury Decide, PostHog Jeeves 9B, AWS Strands Decider 2B, Laya, Respan Span-01, Upstage Solar Decide and Together Tev1.

TL;DR

• Start with TypeSafe Jev. It led or tied both large independent tests, costs $0.042 per million input tokens, and carries most OpenRouter traffic.

• Pick the OpenAI Decisions API for images or OpenAI's compliance terms. $0.10 per million, in beta.

• For long inputs and the lowest bill, Perplexity Decider reads up to 262K tokens per request at $0.02 per million.

• To own the weights, use Cloudflare Clef (open and hosted) or Kev (self-host).

How we ranked them

Each model gets an evidence label. Independent tests means 500 or more cases run by someone with no model in the race: Arize's October 7 test or Benchmark Heaven's JevBench. Vendor claims only means the maker's numbers. Price and usage only (no accuracy numbers) fits none.

Independent tests rank first. Within a label, we order by OpenRouter requests this week, then by Arize accuracy. Anthropic and Google have no native decision model.1 "Where to use it" lists every place found among nine platforms checked October 9 to 11. Only Microsoft-Decision-1 is on a big cloud (Microsoft Foundry); none is on Bedrock or Vertex AI.2 Specs give p50 latency and who measured it ("vendor" marks maker figures), then input limit, inputs, caps and weights.3

The 14 best decision model APIs, ranked

1. TypeSafe Jev

The first decision model, jev-1.13.0 (TypeSafe docs).

• Price: $0.042 per million input tokens.

• Where to use it: TypeSafe's API, the AI SDK, AI Gateway, OpenRouter and Cloudflare Workers AI.4

• Specs: 0.08 s p50 (Arize), 526 ms p99 (agent.nexus); 32K input; text; question cap not published; closed.5

• Evidence: Independent tests. Top or tied in every Arize round, best calibrated, stable when reworded.6

• Best for: Text routers, tool gates and eval judges. See FOMA's Jev how-to.

2. OpenAI Decisions API

POST /v1/decisions on gpt-6-luna, public beta since October 6 (OpenAI guide).

• Price: $0.10 per million, about 2x Jev per call.7

• Where to use it: OpenAI's API, the AI SDK, AI Gateway and OpenRouter.8

• Specs: 0.16 s p50 (Arize), 1,469 ms p99 (agent.nexus); input limit not published; text, images; question cap not published; closed.9

• Evidence: Independent tests. 81.5% at Arize, behind the top three.10

• Best for: Image checks and OpenAI shops. See FOMA's Decisions API vs Jev.

3. Cloudflare Clef and Clef-flash

Open decision models since October 1 (Cloudflare, FOMA).

• Price: $0.24 per million for Clef, $0.09 for Clef-flash on Workers AI (pricing).11

• Where to use it: Workers AI, OpenRouter and Hugging Face.12

• Specs: 0.50 s p50 (Arize); vendor: 209 ms, Clef-flash 39 ms; 64K; text, JSON, images, video; 1–64 questions; open.13

• Evidence: Independent tests. Matched Jev's accuracy at Arize, but cost $0.223 per 1,000 decisions against $0.045.14

• Best for: Open weights with a hosted fallback.

4. Microsoft-Decision-1

Microsoft's hosted Qwen3.5-9B post-train, from October 9 (Microsoft).

• Price: $0.042 per million input tokens; output is free.

• Where to use it: Microsoft Foundry, AI Gateway and OpenRouter.15

• Specs: 0.46 s p50 (Benchmark Heaven); 32K input; text, JSON; question cap not published; closed.16

• Evidence: Independent tests, one so far: sixth of hosted APIs on JevBench, behind Jev on accuracy and calibration.17

• Best for: Azure shops.

5. Liquid d1

Liquid AI's hosted d1 (Liquid).

• Price: $0.04 per million on AI Gateway and OpenRouter.

• Where to use it: Liquid's API, AI Gateway and OpenRouter. Smaller open siblings only.18

• Specs: 0.20 s p50 (Arize); vendor: 200–300 ms; 64K per gateways; text, plus images on Liquid's API; question cap not published; closed.19

• Evidence: Independent tests. 84.8% at Arize at the lowest cloud cost, but the worst wording failure.20

• Best for: Cheap text decisions with controlled wording.

6. Kev

Jared Palmer's open family, 0.8B to 27B (Jared Palmer).

• Price: Free to self-host; Kev 4B is $0.042 on OpenRouter.

• Where to use it: Hugging Face, or Kev 4B on OpenRouter.21

• Specs: 1.03 s p50 for Kev 9B on a laptop (Arize); 64K state (Kev 27B); text; question cap not published; open.22

• Evidence: Independent tests. Kev 9B tied Jev on Arize's routing final but fell to a coin flip when reworded.23

• Best for: Self-hosting and fine-tuning.

7. Perplexity Decider

pplx-decider-v1.1-27b replaced v1 on October 6 at half the price (Perplexity docs).

• Price: $0.02 per million (changelog).

• Where to use it: Perplexity's API, OpenRouter and Hugging Face.24

• Specs: latency not published; 262K; text, JSON, images; 1–128 questions; open.25

• Evidence: Independent tests. Top JevBench capability score, 82.8 against 77.1 for Jev, but untested at full length.26

• Best for: Whole transcripts and contracts.

8. Inception Mercury Decide

Hosted, on Jev's request format, since September 30 (Inception).

• Price: $0.02 per million (50% off $0.04), with a free tier on OpenRouter.27

• Where to use it: Inception's API and OpenRouter.

• Specs: 0.33 s p50 (Benchmark Heaven); 64K per OpenRouter; text, JSON; question cap not published; closed.

• Evidence: Independent tests. Third of 26 hosted APIs on JevBench, with weaker calibration than Jev.28

• Best for: Free experiments.

9. PostHog Jeeves 9B

An open 9B model that reasons before it decides (Hugging Face).

• Price: Free to self-host.

• Where to use it: Hugging Face only.

• Specs: 16.8 s p50 on a laptop (Arize); vendor: 3.3 s on an H100; context not published; text; question cap not published; open.29

• Evidence: Independent tests. 80.3% at Arize, but flipped when reworded.30

• Best for: Offline judging.

10. AWS Strands Decider 2B

AWS Strands Labs' open 2B model from October 1 (Strands).

• Price: Free to self-host.

• Where to use it: Hugging Face and GitHub; an AWS guide self-hosts it on SageMaker AI.31

• Specs: 0.34 s p50 on a laptop (Arize); vendor: about 115 ms; input limit not published; text, experimental images; question cap not published; open.32

• Evidence: Independent tests. 69.0% at Arize, and flipped when reworded.33

• Best for: Laptop experiments and Strands agents.

11. Laya

Convai Innovations' 322M encoder (Arize).

• Price: $0 on AI Gateway (Vercel).

• Where to use it: AI Gateway and Hugging Face.34

• Specs: 0.04 s p50 on a laptop (Arize); vendor: about 35 ms; 8,192 tokens; text; question cap not published; open.35

• Evidence: Independent tests. Last at Arize, 59.8%.

• Best for: Simple, high-volume checks on a CPU.

12. Respan Span-01

Scores agent behaviors you define in plain language (Respan).

• Price: $0.02 per million; Span-01 Lite is free.

• Where to use it: Respan's API and OpenRouter.36

• Specs: latency and context not published; text traces; no cap on behaviors; closed.37

• Evidence: Vendor claims only. Respan reports F1 of 84.3 against 71.5 for Jev on its own benchmark.38

• Best for: Guardrails on agent traces.

13. Upstage Solar Decide

A hosted beta on Solar Mini 4, strong in Korean (Upstage).

• Price: $0.10 per million, $0.05 on OpenRouter during a promotion.

• Where to use it: Upstage's API and OpenRouter.39

• Specs: latency not published; 512K; text; 2–26 options per question; closed.

• Evidence: Vendor claims only. Upstage reports 98.4% on about 400 questions; a small outside test disagrees.40

• Best for: Korean text and very long documents.

14. Together Tev1 4B

An experimental Qwen3.5-4B fine-tune (Together).

• Price: $0.042 per million on OpenRouter.

• Where to use it: Together's API, OpenRouter and Hugging Face.41

• Specs: latency not published; 32.8K window; text; one question, 2–24 options; open weights, licence pending.

• Evidence: Vendor claims only. Together reports 88.0% on its own development set.42

• Best for: Copying the recipe to train your own.

How do the decision model APIs compare?

Which decision model should you use?

• Agent routing and tool gates on text: Jev.

• Screenshots, scans and photos: OpenAI Decisions, or d1, Clef or Perplexity.

• Eval judges: Jev (best calibrated) or Clef.

• Long documents: Perplexity (262K tokens), or Solar Decide (512K, vendor numbers only).

• Data on your own servers: Clef, Kev or Jeeves.

• Already on Azure: Microsoft-Decision-1.

• Lowest bill: Perplexity, Mercury Decide or Liquid d1.

Whichever you pick, set thresholds on your own labelled cases.

So, which decision model API should you use?

For most builders, the answer is Jev: cheap, best tested and easiest to reach. Switch to OpenAI's Decisions API for images or compliance terms. Use Perplexity for long inputs or the lowest bill. Choose Clef or Kev for weights.

Notes

Claude Haiku 5.5 is Anthropic's cheapest, fastest model and GPT-6 Luna is OpenAI's cheapest GPT-6 model; both cost $0.10 per million input tokens and $0.50 per million output tokens on the API, but Haiku thinks out loud for longer and gets 5x more expensive past 100,000 tokens.

TL;DR

• Pick GPT-6 Luna for bulk work (classification, extraction, routing) and for long prompts. It's cheaper per task at every effort level.

• Pick Claude Haiku 5.5 for anything that touches a terminal, a browser or a codebase. It wins Terminal-Bench 4.0 by 33% to 13%.

• Same sticker, different bill: Haiku writes about 3x more output, so Artificial Analysis measures $0.21 per task for Haiku and $0.07 for Luna at max effort.

• Watch the 100K cliff: past 100,000 tokens of prompt, Haiku's price goes 5x.

Anthropic shipped Haiku 5.5 on October 7 and priced it to the cent on top of Luna, which OpenAI launched in September next to GPT-6 Sol (FOMA covered that week). Same sticker. Not the same bill.

What is Claude Haiku 5.5 good for?

The boring, high-volume stuff that used to be too expensive to bother with: summaries, compaction, database queries, classification, live support, browser use, and grunt work as a sub-agent while Opus 5.5 or Sonnet 5.5 does the thinking on a coding job. It's the first Haiku with effort settings (default medium), and it gets a 1M-token context window.1

Anthropic is also refreshingly honest about the ceiling: Sonnet 5.5 and Opus 5.5 "remain better choices" for complex agentic coding. Haiku is the intern you give narrow, well-scoped tickets, not the one you hand the migration.

What is GPT-6 Luna?

OpenAI calls Luna its "most efficient model for focused, high-volume tasks," which is lab for "the cheap one." It also has a context window of about a million tokens, and effort levels from none to max.2 Starting October 8, it's rolling out as the ChatGPT model for Free and Go users, per OpenAI's rollout post (FOMA's issue), so a lot of people's first taste of GPT-6 is this one. OpenAI's Decisions API, the thing that classifies and routes, runs on Luna too (FOMA covered it).

Claude Haiku 5.5 vs GPT-6 Luna: benchmarks

Neutral referee first. Artificial Analysis ran both at max effort. Haiku scores higher on its Intelligence Index (43 to 38), is far better in a terminal (33% to 13% on Terminal-Bench 4.0), wins knowledge work and is about twice as fast. Luna wins SaaS workflows on AutomationBench-AA. Long context is a tie.3

Asterisk on AutomationBench: Artificial Analysis says a pre-release safety issue made Haiku over-refuse, and it plans to re-run the test once Anthropic's fix lands, expecting the score to rise. So Luna's one clean win might be partly Haiku saying no to its homework. Same write-up, better news for Haiku: it makes things up about half as often.4

Anthropic's own table shows even bigger Haiku leads on computer use and coding. Vendor numbers, vendor salt.5

Haiku 5.5 vs Luna 6 pricing: same sticker, different bill

Up to 100,000 tokens, Anthropic's price list matches OpenAI's line for line: $0.10 input and $0.50 output per million tokens, with the same cache and batch discounts. That's 90% below Haiku 4.5.6 Then the price lists stop being twins:

• Claude Haiku 5.5 over 100,000 tokens: 5x on every rate, priced per request by prompt length.

• GPT-6 Luna over 272,000 input tokens: 2x input and 1.5x output.7

Anthropic says prompts up to 100,000 tokens were around 90% of requests to the previous Haiku, so most calls never hit the cliff. The sneakier cost is tokens. Artificial Analysis measured Haiku at max effort writing roughly 3x Luna's output per task, and reasoning tokens bill as output.8 Same price per token, more tokens. Here's the math.

The per-task cost formula

Cost per call = (input tokens × input price + cached tokens × cache price + output tokens × output price) ÷ 1,000,000.

Assumptions below: no caching, standard tier, list prices as of October 8, 2026, and made-up but typical token counts. Swap in your own logs; your logs know things I don't.

1. Extraction, 3,000 tokens in, 300 out. $450 per million documents on either model. If Haiku writes 3x the output, it's $750. The gap is 1.67x, not 3x, because input is two-thirds of the bill.9

2. Long document, 150,000 tokens in, 2,000 out. Luna: $0.016. Haiku crosses 100K: $0.080. FIVE TIMES the price before Haiku writes a single extra token.

3. Huge document, 300,000 tokens in, 2,000 out. Luna crosses 272K: $0.0615. Haiku: $0.155. About 2.5x, because Luna finally has a surcharge of its own.10

Now the measured version. Artificial Analysis puts cost per task at $0.21 for Haiku and $0.07 for Luna at max effort, and Luna wins every effort label (release comparison).11 Line them up by score instead and it's basically a tie on price, with Haiku answering much sooner.12 One more hedge: those Haiku figures don't include the over-100K surcharge yet, so they run low.

Which one should you use?

• High-volume classification and extraction: Luna. Cheaper per task at every effort level, a much faster first token at low effort, and the Decisions API is sitting right there.13 Switch to Haiku when a confident wrong answer costs you more than the tokens do.

• Agents and tool use: Split. For SaaS and API workflows, Luna for now on AutomationBench, until the re-run. For terminal, browser and computer use, Haiku, and it isn't close (33% versus 13% on Terminal-Bench 4.0).

• Long context: Luna. Same long-context score, and it stays at base price until 272K while Haiku goes 5x at 100K.

• Coding sub-agents: Haiku. It's what Anthropic built it for, and the Terminal-Bench gap is wide. Keep each sub-agent's prompt under 100K and the sticker price holds. Set effort on purpose: max costs far more tokens than xhigh for a small gain.14

• Knowledge work and document analysis: Haiku, on GDPval-AA.

Haiku 5.5 vs Gemini 3.8 Flash

Gemini 3.8 Flash is the third option, and it is not priced like the other two: $0.75 input and $3.75 output per million on a launch rate, which is 7.5x Luna's sticker, and the price doubles in January (Google's pricing).15 Artificial Analysis scores it 41, between Luna and Haiku, with the best factual recall of the three.16 Pick it if you already live on Google's stack or need recall. At 7.5x the token price for a score between the other two, it's hard to justify on cost alone.

What else is cheap?

• Claude Sonnet 5.5: $2 / $10, and on Haiku's launch day Anthropic halved its cache reads to $0.10, which it says makes Sonnet around 20% cheaper on most agentic work. That's the real step up when the coding gets hard. FOMA's Sonnet 5.5 issue has the rest.

• Going bigger: FOMA's GPT-6 Astra vs Claude Fable 5.1 comparison covers the expensive end of both lineups.

So, Haiku 5.5 or Luna?

You opened this to find out which cheap model to use, and the price tag won't decide it because it's identical. Use Luna for bulk classification, extraction and long documents: it costs less per task and has no cliff at 100K tokens. Use Haiku for coding sub-agents, terminals and browsers: it's the stronger agent and earns its extra tokens there. If you run both kinds of work, use both and route each job to the right one.

Notes

OpenAI's Decisions API is an endpoint that answers a fixed question about text or images with a probability, a choice from your list, or a score, instead of generated text; it entered public beta on October 6, 2026, runs only on GPT-6 Luna, and costs $0.10 per million input tokens with no output charge.

TL;DR

• Pick Jev for text-only classifiers, routers and guardrails. In the one independent test so far, it matched the Decisions API on accuracy at about half the cost per call, with a much shorter slow tail.

• Pick the Decisions API when the input is a screenshot, scan or photo, or when you need OpenAI's zero data retention, HIPAA or EU data residency. Jev reads text only.

• Price: $0.10 per million input tokens for Decisions against $0.042 for Jev. Neither charges for output.

• Status: public beta, with general availability expected "in the coming weeks," per OpenAI.

OpenAI first showed the Decisions API at DevDay on September 29 as a limited preview (FOMA covered it that week). A week later the changelog opened it to everyone. It's the same shape as TypeSafe's Jev, which shipped two weeks before DevDay: you write the question and the allowed answers, the model picks. No blank page.

What is the OpenAI Decisions API?

You send POST /v1/decisions with an input (text, images or both) and a list of questions. Each question has one of three types (OpenAI's guide):

• Predicate: "is this true?" Returns a probability from 0 to 1.

• Choice: "which of these?" Returns one of the values you supplied, plus the probabilities and a confidence field.

• Score: "how bad, on this scale?" Returns a probability-weighted average of the levels, so it can land between them.1

You can ask several independent questions about one input in one request. If question two depends on the answer to question one, that's two requests. An answer can also come back as a refusal, so your code needs that branch.

OpenAI says it is about 10x faster than asking GPT-6 Luna the same thing through the Responses API. That's a comparison with OpenAI's own chat endpoint, not with other decision models.2

OpenAI Decisions API pricing and availability

Here's the price sheet, as of October 9, 2026:

• Decisions API (gpt-6-luna): $0.10 per million input tokens. No output, cache-read or cache-write charges (guide). Regional processing and long-context multipliers apply.3

• GPT-6 Luna through the Responses API: $0.10 input and $0.50 output per million tokens (model page).

• TypeSafe Jev: $0.042 per million input tokens, output free (TypeSafe's model docs).

• Status: Decisions is a public beta. Jev launched on September 15.

• Inputs: Decisions takes text and images. Jev takes text only.4

• Data controls: Decisions supports zero data retention and HIPAA for eligible customers, with data residency in the US and Europe.

OpenAI Decisions API cost per call

The sticker is 2.4x Jev per token, but the bill depends on how many tokens each vendor counts. The only head-to-head so far, a 9,000-call test by agent.nexus published October 7, measured $0.22 per 1,000 calls on Decisions against $0.11 on Jev. So about 2x.5

The catch is per-question overhead. In that test, each extra yes-or-no question added about 144 billed tokens on Decisions and almost nothing on Jev. One question per call costs 1.5x Jev. Eight questions costs 6.8x.6 If your design packs many checks into one call, which is the point of these models, price it with your real question count.

Is the Decisions API on Azure or Bedrock?

Not that I could find. The guide documents only OpenAI's own endpoint. A Microsoft moderator said on October 7 that there is no confirmed date for Microsoft Foundry (Microsoft Q&A). Vercel's AI Gateway did add it on October 7, under the model ID openai/gpt-6-luna-decisions (Vercel changelog).

OpenAI Decisions API vs Jev

Same job, different trade-offs. Ties on the average call, loses on the slow ones. In the agent.nexus test, on eight real text checks:

• Accuracy: a tie. 233 errors for Decisions and 221 for Jev on the same 2,155 labelled cases, inside each other's error bars.

• Typical speed: a tie. Median 268 ms against 269 ms.

• Slow calls: Jev. At the 99th percentile, Decisions took 1,469 ms against Jev's 526 ms.7

• Cost per call: Jev, at about half the price.

Two big asterisks. The testers run Jev in their own product and say so, and they tuned thresholds on Jev's outputs. And they tested text only. Images are where Decisions has no Jev rival at all: "is this a login page," "is there a card number in this upload," "is this scan signed." Jev can't take the input.

Jev also has more neighbours now. Cloudflare's open-weight Clef models do the same job and return a probability instead of a paragraph (FOMA's October 8 issue).

How to try the Decisions API

1. Update your SDK. The guide lists Python 3.26.0 and JavaScript 7.30.0 as the minimums, with Go, Ruby and Java versions too.8

2. Try a question in the Decisions playground before writing code.

3. Send images as base64 data URLs. Hosted image URLs and file IDs aren't supported in the beta.

4. Label 100 to 200 of your own cases and set thresholds from them. The guide says the same: pick thresholds from the cost of false positives and false negatives.

If you already call Jev, FOMA's Jev how-to shows the kind of checks worth moving, and the Vercel gateway endpoint serves both models with the same request shape.

Who is the Decisions API for?

Teams already on OpenAI who need a fast yes, no or pick-one inside an app: support routing, content moderation, agent guardrails. Teams whose inputs are images. And teams whose compliance list includes zero retention or HIPAA, where a new vendor contract is the slow part.

It's not the cheap option for text. Jev is cheaper per token and, in the one test we have, steadier at the tail. TypeSafe's story is in FOMA's Jev issue.

So, Decisions API or Jev?

You want to know which decision model to call, and for plain text the answer is Jev: same accuracy in the only independent test, about half the cost per call, fewer slow calls. Pick the Decisions API when your input includes images, when your compliance team needs OpenAI's data controls, or when adding a vendor costs more than the token bill. It's a beta, so re-run your own cases when it reaches general availability.

Notes

TypeSafe Jev is a System One decision model you drop into an agent harness to answer closed questions with probabilities, so your code can route, gate, compact, and verify without asking a chat model to write a paragraph about it.

TypeSafe launched Jev on September 15, 2026. Founder Diogo Almeida helped build the RLHF stack behind InstructGPT and ChatGPT, then shipped a flagship model that cannot write a sentence. FOMA covered the launch the week it shipped. This page is not another "what is Jev" explainer. You already have (or are building) an agent loop. The question is what Jev does inside that loop.

What can you do with Jev in an agent harness?

An agent harness is the boring machinery around the LLM: ingress, tools, memory, safety, egress. Most of those steps are judgments with a small answer space. Chat models keep answering them with essays. Jev answers them with typed decisions your code can threshold.

Per the System One docs and API, every question is one of three primitives:

• Choice: pick one option from a set you define (up to 255). You get the winner, a full probability distribution, and confidence.

• Score: rate against ordered levels (at least two, up to 10). You get a probability-weighted score that can land between levels, plus confidence.

• Noul: yes/no. You get a probability from 0 to 1 that the answer is yes.

You send a state (string, JSON, or text array) plus a map of questions to POST https://api.typesafe.ai/v1/systemone. Sampling is parallel across questions, not one token at a time. TypeSafe's launch post quotes end-to-end latency of 70ms to 500ms, input at $0.042 per million tokens, and free output. The models page for Jev 1.13 (jev-1.13.0) lists a 64k token request budget and 32k for state plus the longest question. Text only. No images yet.

That shape is the product: unstructured agent state in, typed probabilistic decisions out. Your generative model still writes, plans, and calls tools. Jev owns the fuzzy branch.

Where does Jev sit in the agent loop?

Think of the harness as a chain of gates. Each gate is a closed question. Here is how builders actually wire it, mapped to the primitives:

1. Ingress / model router → Choice. Commit the request to Fast vs Deep LLM (or cheap vs expensive, local vs frontier). One call, one option set, no "let me think about which model…" essay.

2. Pre-execution / safety gate → Noul or Choice. Halt for human approval if the planned action looks risky (delete, spend money, leave the sandbox). Noul for a single risk check; Choice when you need auto-allow / ask-human / hard-deny.

3. Data ingestion / injection filter → Noul. Drop or sanitize payload text that looks like a prompt injection before it enters the context window. State is data; TypeSafe's jaggedness page is blunt that adversarial framing can move answers, so write precise criteria and test edge cases.

4. Runtime maintenance / context compactor → Noul. Ask whether the transcript is bloated enough to trigger a summarization sweep. If yes, your code runs the summarizer (a generative model). Jev only fires the gate.

5. Egress / output verifier → Score. Rate the draft against a rubric (grounded, complete, on-policy). Re-run the loop if the score fails your threshold; ship if it clears.

Those five are the lifecycle Adir keeps pointing at: router, safety, filter, compact, verify. Tool selection sits next to the router. Give Choice a list of allowed tools for this turn (search, shell, browser, none) and let probabilities drive which one your harness exposes. High confidence and a clear winner: call it. Low confidence: ask the human or fall back to a smaller tool set.

You can fan several of these into one System One request against the same state. That is the whole point of parallel sampling. A ticket-router demo in the quickstart already mixes Choice + Score + Noul on one message. An agent step does the same: route model, gate risk, pick tool, score the last output, in a single round trip.

How do you use TypeSafe Jev for agent decisions?

The quickstart is short on purpose.

1. Get an API key from the TypeSafe dashboard.

2. Call POST /v1/systemone with jev-latest or a pinned jev-1.13.0.

3. Put the current harness state in state: user goal, last tool result, policy snippets, transcript slice.

4. Ask only the closed questions your code can act on.

5. Branch on probability and confidence. Do not ask Jev to explain itself.

Official clients: Python typesafe-sdk and JavaScript @typesafe-ai/sdk. There is an agent skill at github.com/typesafe-ai/skills for Claude Code, Codex, and friends. Weights are not open. SDKs and skills shipped; a downloadable Jev checkpoint did not.

Keep the state thin. The jaggedness docs say accuracy falls when you dump irrelevant junk into state. Filter in code first. Keep math, counting, and date arithmetic out of the model. Do not chain Choices to fake generation. If you need a paragraph, call a real LLM.

What are other people doing with decision models?

A week after launch, FOMA noted the sequel: Jev never open-sourced, the launch video pulled 36 million views, and the internet shipped clones anyway (Laya, OpenJev, Bespoke Nimble, Kev-0.5B on a MacBook), with Braintrust wiring Jev into evals at roughly 400x cheaper scoring. The category landed even when the proprietary model stayed behind TypeSafe's API.

OpenAI later shipped a similar product shape. The Decisions API on GPT-6 Luna (POST /v1/decisions) takes shared input plus typed questions (predicate, choice, score) and returns probabilities. FOMA flagged it around DevDay as the same idea Jev already shipped: judgment without a blank page. OpenAI's docs put Decisions at about 10x faster than the Responses API, with pricing at $0.10 per million input tokens and no output charge on that endpoint. Decisions also accepts images. Jev, as of the models page, is text only. Same category. Different vendor, price, and modality.

TypeSafe's own workflow evals claim 193.6x faster / 444.6x cheaper against an average of GPT-6 Astra and Claude Fable 5.1, with bias caveats attached. Treat those as vendor receipts, not an independent bake-off. The useful signal for harness builders is the shape: many independent questions, fine-grained probabilities, discrete branching in your code.

When should you not use TypeSafe Jev?

Skip it when the answer space is open.

• Generation: drafts, emails, code patches, chain-of-thought explanations. Jev is not trained to write. The jaggedness page says forcing generation by chaining choices is slow and bad.

• Open-ended planning: inventing a multi-step plan you have not enumerated. Choice needs the options in advance. If you do not know the option set, you need a generative model (or a planner that proposes options, then Jev ranks them).

• Exact math, counting, date ordering: keep that in code. Jev is not a calculator.

• Image / audio / video judgments: not supported on Jev 1.13. Use OpenAI Decisions or a vision model for that lane.

• Huge unfiltered transcripts: filter first. Context rot is real on this model class.

Use it when the call sits inside software and the answer is closed: classify, route, score, guardrail, re-rank, verify a citation against a source, gate a tool call, decide whether to compact. Pair it with a real LLM. Do not ask it to write your emails.

FOMA's take: the agent race keeps shipping bigger chat models into loops that mostly need a traffic light. Jev is the rare launch that makes every classify/route/score hop feel like infrastructure. If you already built the harness, that is the missing piece. What is FOMA.

Nano Banana 2 is Google's general-purpose image model (API name Gemini 3.1 Flash Image), and DALL-E is the OpenAI image brand people still type after OpenAI retired it and moved ChatGPT image generation to ChatGPT Images.

People are still searching for a fight with a retired brand. "dall e vs nano banana" pairs a February 2026 Google model with a product name already off OpenAI's API. The live opponent is ChatGPT Images 2.5.

What is Nano Banana 2?

Google shipped Nano Banana 2 on February 26, 2026. Silly name, real model: photoreal lighting, instructions it follows, and Gemini's world knowledge, including live web search.

Google's launch specs: resolutions from 512px to 4K, consistency for up to five characters and up to 14 objects, legible text you can also translate in place, plus SynthID and C2PA. Google said the Gemini app's SynthID check had been used over 20 million times since November 2025.

FOMA's March 5 issue reported it at #1 on Arena and Artificial Analysis text-to-image that week. Treat that rank as launch week only.

The rest of the family, from the June 30 Lite post and the pricing page: original Nano Banana (Gemini 2.5 Flash Image) has a listed shutdown of October 2, 2026. Nano Banana Pro (Gemini 3 Pro Image) stays the accuracy model. Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image, June 30, 2026) is the fast cheap one. Google has not announced a Nano Banana 3.

Is DALL-E still the image model in ChatGPT?

DALL-E 3 was OpenAI's image model in ChatGPT, later sold as dall-e-3. On November 14, 2025, OpenAI told developers that dall-e-2 and dall-e-3 would leave the API on May 12, 2026. The deprecations page still lists that date. The replacements named then were gpt-image-2, gpt-image-1, or gpt-image-1-mini. The DALL-E 3 model page now says the model has been removed and recommends GPT-Image-2.5 Sunburst. You can still type DALL-E. The API will not answer to it.

In the app, OpenAI's ChatGPT release notes set August 30, 2026 as the retirement of the official DALL-E GPT, and pointed people at ChatGPT Images.

ChatGPT Images 2.5 shipped September 8, 2026, measured against Images 2.0 and the API model GPT-Image-2. OpenAI said people were already making more than 3 billion images a week across ChatGPT Images and the GPT-Image API. FOMA covered it that week. In the API it splits into gpt-image-2.5-flare (fast) and gpt-image-2.5-sunburst (tighter edits, longer waits).

Which one looks better, and which one renders text?

Nano Banana 2 is photoreal at Flash speed, with web search for a real subject and a resolution you pick, up to 4K. Google leads the announcement with accurate, legible text, including translation of text already in the image. A sign, a headline, or a poster in another language is this model's job.

ChatGPT Images 2.5 aims at natural lighting, people who stay recognizable from a reference photo, real-world information in the picture, and layouts that include transparent backgrounds. It also edits a piece of copy while the subject and brand treatment stay put. For a photo of a real person, start there.

Which one is better at editing?

Images 2.5 is built to change only what you asked, including on messy subjects, and to keep earlier edits alive across a long chat. You talk to the picture: Sketch (type @Sketch and draw the layout), comments on the image, templates such as Poster and Merch, and shared prompts someone else can rerun on their own photos.

Nano Banana 2's edit story is fast iteration, stricter instruction following, and that five-character / 14-object consistency. Google kept Nano Banana Pro for "maximum factual accuracy." A thread of "change the background, keep her face" is an Images 2.5 job. The same character across six print-resolution frames is a Nano Banana 2 job.

Which one is faster?

Images 2.5 cuts latency by up to 50% versus Images 2.0. Flare is the fast API default: higher quality than GPT-Image-2 at 50% lower latency, per OpenAI. Sunburst waits longer for a tighter edit. The published stopwatch time on Google's side is for Nano Banana 2 Lite, the cheap one: 4 seconds for text-to-image.

How much does Nano Banana 2 cost vs ChatGPT Images?

Google prints a per-image price. OpenAI prints a per-token price, and says the GPT Image 2 calculator will not estimate 2.5.

Standard rates for gemini-3.1-flash-image, from Google's pricing page. The free API tier is unavailable.

• $0.045 per 0.5K image, $0.067 per 1K, $0.101 per 2K, $0.151 per 4K. Input is $0.50 per million tokens, so reference images add to the picture price.

• Batch is $0.034 per 1K. Search grounding is 5,000 free requests a month, shared across Gemini 3, then $14 per 1,000.

At 1K, $0.067 is half of Nano Banana Pro's $0.134 per 1K or 2K image. At 4K, Nano Banana 2 is $0.151 and Pro is $0.24. Nano Banana 2 Lite is $0.0336 per 1K (the June post rounded that to $0.034).

GPT-Image-2.5 Flare bills image output at $30 per million tokens ($15 on Batch), text input at $5 per million, and image input at $8 per million. Quality runs low through max. The free API tier is unsupported. OpenAI says those rates match GPT Image 2, and that there is no 2.5 token estimate in the old calculator, so there is no official dollars-per-picture number. Flare's $30 per million is a lower rate than Nano Banana 2's $60 per million image-output tokens. OpenAI has not published how many tokens a 2.5 image spends, so the lower rate can still cost more per picture.

In the apps, Images 2.5 is on every ChatGPT tier, plus ChatGPT Work and Codex. Neither launch post states a free-app image cap.

Where can you use each one?

Google's February rollout put Nano Banana 2 in the Gemini app (Google said it would replace Pro on Fast, Thinking, and Pro, with Pro and Ultra subscribers still able to regenerate on Pro from the three-dot menu), Search AI Mode and Lens (141 new countries and territories, eight more languages), AI Studio, the Gemini API, Vertex AI (preview at launch), Antigravity, Flow (default, zero credits), and Google Ads. Lite was also headed to NotebookLM, Google Photos, and Stitch. Check the picker before you promise a client which Banana they will get.

Images 2.5 is on ChatGPT, ChatGPT Work, and Codex, all tiers, desktop, mobile, and web, with Flare and Sunburst in the API. Adobe said the 2.5 models were already in Firefly. Google watermarks with SynthID and C2PA. OpenAI uses C2PA, an invisible watermark, and checks on prompts and outputs.

Should you use Nano Banana 2 or ChatGPT Images?

• Posters, signs, localized text, 4K, search-grounded subjects, a character held across frames: Nano Banana 2.

• High-volume drafts at 4 seconds and about $0.034 a 1K image: Nano Banana 2 Lite.

• The accuracy job Google holds back: Nano Banana Pro.

• A chat where you keep editing one picture, a doodle as the layout, or a reference photo of a real person: ChatGPT Images 2.5. Flare for the fast API default. Sunburst when the edit has to be exact.

• A per-image API price you can put in a spreadsheet tonight: Google. OpenAI's 2.5 bill depends on a token count the docs will not estimate.

March's "crushes DALL-E" line was about a model that has since left the API. Against ChatGPT Images 2.5 the split is: Google on published price, in-image text, resolution, and Search; OpenAI on the editing loop already open in ChatGPT.

Is there a Nano Banana 2 jailbreak?

People keep searching for one. This page will not publish a bypass, a prompt, or a filter workaround. Google marks Nano Banana images with SynthID and C2PA. OpenAI checks prompts and images on ChatGPT Images 2.5 and watermarks the output. A prompt that slips those checks breaks the product rules. If the model refuses a picture, rewrite the request so it stays inside the rules.

What else can you use?

• Qwen-Image-2.1 is Alibaba's open-weight generate-and-edit model: a 7B visual generator, up to 10 reference images, native transparent RGBA, under the Qwen Research License. FOMA's September 23 issue reported early testers running it on a 32GB Mac and claiming it beat Nano Banana 2 on several benchmarks. That claim is from testers.

• Muse Image, in FOMA's July 9 issue, was reported at #2 on Image Arena at launch, behind GPT Image 2. The later Muse personal agent is a different product.

• Luma Uni-1, in FOMA's March 26 issue, led human-preference Elo in several categories that week. The March 12 issue reported it ahead of Nano Banana 2 and GPT-Image 1.5 on logic tests. Both notes predate Images 2.5.

• Midjourney announced a medical body scanner in June, which puts it outside this comparison.

FOMA covered these the weeks they shipped. What is FOMA? is the short version of the newsletter.

A personal AI agent is an assistant that works for you across days, not just one chat: it has a computer or browser of its own, keeps going after you close the app, and asks before it spends money or sends anything.

We compared 13, split by one question: do you want one agent or a team? Four left the chat box in one week (FOMA, October 1).

TL;DR

• For errands (email, calendar, shopping): Muse is the best free personal AI agent and tops Assistant Benchmark at 9.3, but only in the US and Canada. Elsewhere, start with Gemini Spark if you pay for Google AI Pro ($19.99).

• For projects that need a team of agents: Grok Bot is the only AI agent team on general release, from $20 a month on Cursor Pro. One tester says plan on a $100 tier.

• Is ChatGPT Dots worth it? Only if you already pay for ChatGPT Pro ($100 and up). It is one agent, and Pro doesn't offer it in the EEA, UK or Switzerland.

• Manus Cue is the closest Grok Bot rival, but invite-only.

One agent or a team?

Type A, personal assistant: one agent for errands (email, calendar, reminders, shopping, bookings) in WhatsApp, iMessage or an app.

Type B, AI agent team: a multi-agent team is several agents, or teammates, that share a computer (browser and terminal), code and git, and stored logins they use without seeing passwords; they run on schedules or event triggers and message each other to hand off work. Below: agent teams, also called agent swarms or an AI workforce. Cue says "personal agents" but gives you several.

Hybrids have one agent with team-style tools (a cloud computer, terminal or coding hand-offs): Dots and Genspark Claw.

How we ranked them

Independent tests means a reviewer with no product in the race ran it: Assistant Benchmark (15 personal tasks) or Pat Simmons at AI for Mortals (three identical jobs, October 6). Vendor claims only means the maker's docs, checked October 11; "not published" means they are silent.1

Disclosure: FOMA's research is assisted by Grok Bot, one of the products ranked here. Its entry uses only Cursor's docs and independent tests.

Best personal assistants (type A), ranked

1. Meta Muse

Type A. Meta's agent, in its app or WhatsApp.

• Price: Free; Power $20 a month; Maximum $100. Paid tiers buy bigger weekly allowances; run-out behaviour: not published.2

• What you get: One agent on its own cloud computer; a separate Sentinel agent gates internet access. Uses stored logins unseen; one-time Stripe Link cards. Team agents, agent-to-agent and schedules: not published.

• Availability: US and Canada, in limited testing.3

• Evidence: Independent tests. First on Assistant Benchmark at 9.3, best at purchasing and email, but a D on AI for Mortals' coffee-shop job.4

2. Hark Pro

Type A. One persistent chat, on web, iOS and Android since October 6 (Hark).

• Price: Free (the Hark Pro plan); Pro² $20 a month for 2x usage; Pro³ $100 for 10x. Same features on every plan.5

• What you get: One agent by design ("no juggling agents"). Handoff cloud computer, encrypted vault for cards and logins, scheduled tasks, repeatable workflows, Google Workspace and Microsoft 365. Code and multi-agent: not published.

• Availability: Region limits not published.

• Evidence: Vendor claims only.6

3. Gemini Spark

Type A. Google's agent in the Gemini app (Google help).

• Price: Included in Google AI Pro ($19.99 a month, ₪74.90 in Israel) and Ultra (from $99.99). Run-out behaviour: not published.7

• What you get: One agent with schedules and Workspace actions. No computer of its own: it drives your Chrome or a remote browser that stops at sign-ins. Can use Google Password Manager logins.

• Availability: Gemini countries except the EEA, UK, Switzerland and Nigeria; Israel is in.

• Evidence: Vendor claims only. See FOMA's Spark launch coverage.

4. Poke

Type A. A text-message assistant in iMessage, Telegram, WhatsApp and RCS (Poke docs).

• Price: Free, Pro $19 and Ultra $199 a month; yearly costs less.8

• What you get: One agent for email, calendar, reminders, web search and integrations. Computer, vault and multi-agent: not published.

• Availability: Worldwide.

• Evidence: Independent tests. 6.3 on Assistant Benchmark, best at personality.

5. ChatGPT Dots (hybrid)

Hybrid. OpenAI's always-on agent (OpenAI docs).

• Price: ChatGPT Pro at $100, $200 or $500 a month, or a Business Premium seat; higher tiers only add usage.9

• What you get: One dot for now. Its own cloud computer plus one of yours; hands coding to Codex, which can push to GitHub; saves sign-ins to Passwords; schedules; Slack, Teams and calls.

• Availability: Pro excludes the EEA, UK and Switzerland; Enterprise is admin-enabled.

• Evidence: Independent tests. 8.4 on Assistant Benchmark; B on AI for Mortals' coffee-shop job.10

Betas: Haven, Instinct, Tab

Haven is a free memory-first assistant on calls, text and email; no paid tiers published.11

Instinct is an agent you text or call that "can use a phone and computer"; waitlist, price not published, 8.5 on Assistant Benchmark.12

Tab has its own number, computer and wallet over iMessage and WhatsApp; waitlist, price not published, 5.9.

Best AI agent teams (type B), ranked

1. Grok Bot

Type B. xAI's bots, sold through Cursor and SuperGrok (Cursor docs, FOMA). See the disclosure above.

• Price: Cursor Pro $20, Pro+ $60, Ultra $200 or Teams from $40 a seat; SuperGrok $30, Plus $100 and Heavy $300 (X Premium+ included) match Pro, Pro+ and Ultra. Plans don't stack; weekly usage, then optional on-demand billing. SuperGrok Lite excludes it.13

• What you get: Unlimited bots (teammates) sharing one persistent cloud computer (browser, files, terminal), or running on your own machine. Bots message each other and hand off work; group chats hold 2 to 6 bots. Routines run on schedules or events (Slack, GitHub, Linear, email, webhooks and more). Masked secrets it fills but never reads; Team Bots. 1Password and an explicit git workflow: not published.14

• Availability: General release; region limits not published.

• Evidence: Independent tests. Won all three AI for Mortals jobs, but 7.3 on Assistant Benchmark, which grades personal errands.15

2. Manus Cue (invite-only)

Type B. Manus's app for personal agents.

• Price: Free in early access with an invite code; a US phone number costs extra ($9.99 in its iOS app).16

• What you get: Several teammates, each with its own email, phone number, wallet and computer. In a group chat with a shared goal, "they hand work to each other". Code, vault and event triggers: not published.

• Availability: Early access on web, desktop, Android and iOS.

• Evidence: Vendor claims only.

3. Genspark Claw (hybrid)

Hybrid. One agent on a dedicated cloud computer, or local on your Mac or PC (Genspark help).

• Price: Cloud Computer Lite $9.99, Standard $19.99 or Powerful $39.99 a month, plus Genspark credits; local mode skips the subscription. Out of credits: top up or upgrade.17

• What you get: Terminal, files, GitHub, schedules and a Heartbeat check; logins saved in its browser; sub-agents; joins WhatsApp, Slack or Teams groups. One per person: "Claw is never shared".

• Availability: Region limits not published.

• Evidence: Vendor claims only.

Also compared, and what's out of scope

Lindy is a work assistant: one per person plus a shared Slack "Teammate", trigger-based routines and browser computer use from Pro. It costs from $49.99 per user a month (a $29.99 Plus also appears, likely a price test) to $199.99 for Max, after a 7-day trial with $50 in credits.18

Perplexity Personal Computer runs its Computer agent on your own Mac or PC with scheduled work, on Pro (from $17) or Max ($200, 10,000 credits a month).19

Out of scope: Claude Cowork (a hand-off tool) and Sesame (a voice companion).

Value verdict: which type is worth paying for?

Personal assistants: Muse is the value pick in the US or Canada: free, and top of the broadest independent test. Elsewhere, Spark is the cheapest add-on for Google AI Pro payers; Poke at $19 works anywhere. Hark is untested. Dots is overpriced for errands alone.

Agent teams: Grok Bot is the only one on general release and won AI for Mortals' tests; budget $60 to $100 a month for daily use. Cue may compete once it opens. Claw suits you if one agent with a terminal is enough.

So, which personal agent should you pay for?

If you pay for Google AI Pro, try Spark first. If you pay for ChatGPT Pro, your dot is included, so use it. If you pay for Cursor, SuperGrok or X Premium+, Grok Bot is already in your plan, and it is the pick for projects that need a team of agents. For errands, start on Muse's free tier in the US or Canada, or Spark or Poke elsewhere. Pay for a $100 tier only when limits block real work.

Notes

Hey everyone, FOMA ice cream guy is back with some more AI news. Google shipped a frontier model to cyber defenders first, with the guardrails off. Then OpenAI put GPT-6 in ChatGPT for everyone, and the answer shows up with buttons. Anywayyyyyy, here are the top stories.

Google’s Gemini 4 Argon goes to cyber defenders first, guardrails off

Google’s new frontier model, Gemini 4 Argon, is not in the app. It goes first to trusted cyber defenders in the Fairwind program, and for them it ships without cyber guardrails. Output goes from 64K tokens to 1 million, at an intro price of $2 / $10 per million, then $4 / $20. Wiz already used it to find a critical hole in hospital software that earlier models missed.

read more →

ChatGPT’s Intelligent UI answers with a diagram you can tap

OpenAI trained GPT-6 to decide when a reply should be a picture, a chart, a form, or buttons, instead of another paragraph. Ask about a bicycle and you get a diagram you can poke. Ask it to split a bill and it builds the tool in the chat. Plus, Pro, Business, and Enterprise get Sol today. Free and Go get Luna tomorrow.

read more →

Mistral previewed Large 4 and named it Le Chonk

Mistral put Large 4 on a public API preview and, very officially, nicknamed it Le Chonk. 1 trillion parameters, 49 billion active, text and images in, text out. Weights are supposed to drop by the end of the month, but the reinforcement-learning run is still going, so the checkpoint you can call today is not the one they plan to publish.

read more →

Reflection’s Beam is a 501B open model that only runs 23B

Reflection announced Beam, a 501 billion parameter open model that activates 23 billion at a time, for coding and agent work. Apache 2.0 weights are due this month. Reflection says it matches GLM 5.2 on hard reasoning at 3 to 4 times less compute, and that comparison is theirs.

read more →

Cloudflare’s Clef models return a probability, not a paragraph

Cloudflare shipped Clef and Clef-flash, open-weight models that return a typed probability instead of a paragraph. Hosted on Workers AI, or download the Apache 2.0 weights. Clef-flash is post-trained from Qwen3.5-9B.

read more →

GPT-6.1 Sol gets an ultrafast tier at about 300 tokens a second

OpenAI added an ultrafast tier on GPT-6.1 Sol, rated around 300 tokens a second. The model was already last week’s story.

read more →

OpenAI fired three safety researchers over an alleged leak

The Wall Street Journal reported OpenAI fired three safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, for allegedly sharing confidential information with an outside AI-safety group. OpenAI says they mishandled sensitive information outside company procedure. The same week David Robinson, who wrote the safety reports for 12 launches over three and a half years, quit and told The Atlantic the culture is broken.

read more →

EmbeddingGemma 2 files video and audio in the same drawer

Google released EmbeddingGemma 2, 740 million parameters, Apache 2.0, mapping text, images, video, and audio into one space. The text-only slice is 270 million parameters, small enough to run on a laptop. A voice memo can search a video.

read more →

FLUX 3 edits one box of a picture, and Nano Banana 2.1 got cheaper

Black Forest Labs released FLUX 3 Image: up to 4K, up to ten reference images, and edits that are supposed to leave the rest of the picture alone. Open weights are promised in the coming weeks. Google rolled Nano Banana 2.1 through Gemini, AI Studio, Search, and Ads at $0.034 a picture, versus $0.134 for the previous Pro image model.

read more →

OpenAI posted 722 math papers from a model you cannot use

OpenAI published 722 math manuscripts, grouped into 372 families, from an unreleased internal model. The repo is Apache 2.0, and 162 have a formalized main result. The rest might be wrong, none of it is peer reviewed, and the model that wrote them is not for sale.

read more →

Earendil shipped Pi 1.0, the harness people already run

Earendil shipped Pi 1.0, the stable cut of the coding harness. New: Codemode, virtual models, deferred tool loading, and Anthropic cache warming. Pi Durable ships beside it, an experimental layer for jobs that keep running after you close the laptop.

read more →

Microsoft’s streaming transcriber beat 37 other models

Microsoft’s MAI-Transcribe-2-Streaming ranked first of 38 models on final-transcript accuracy, at 2.5% word error. Microsoft also says it is 55% faster and 60% cheaper than ElevenLabs. Those last two numbers are Microsoft’s.

read more →

Wispr Flow’s notetaker spells the names on your calendar

Wispr Flow’s Notetaker sits in the meeting and uses your dictionary and calendar, so product names, acronyms, and odd names come out spelled right. The notes go into Claude or ChatGPT through a connector.

read more →

ChatGPT will watermark EU text inside the word choice

OpenAI’s textGrain watermark is coming to eligible ChatGPT and Codex text in the EU over the coming weeks. It nudges which words the model picks, so there is nothing to see, and the API switch is opt-in and off by default. The detector starts with approved researchers, and no watermark does not prove a person wrote it.

read more →

Tavus Griffin talked on video, and 48% of people thought it was human

Tavus launched Griffin, a video model for live face-to-face calls. On those calls, 48% of people who talked to it thought it was a real person. That figure is Tavus’s, from their own page.

read more →

The White House told agencies to stop saying AI

Trump signed an order telling the executive branch to say Super Intelligence, or SI, instead of AI. For now, SI means whatever the statute already calls artificial intelligence. Old regulations and contracts can keep the old word.

read more →

ChatGPT Hermes is the unreleased name for OpenAI's workspace agents: shared coworkers inside ChatGPT that run on a schedule, connect to your team's tools, and keep working after you close the tab. OpenAI never shipped a product called Hermes. TestingCatalog found that name in the ChatGPT web app on April 6 and April 21, 2026, on an agent studio framed as always-on teammates. On April 22, OpenAI published it as workspace agents.

You searched "chatgpt hermes." The thing that exists is workspace agents. Plus does not get it. Free, Go, and Pro don't either. Search "hermes agent" and you may mean Nous Research's open-source agent, covered below. Nous doesn't say whether a Plus login works with its agent either.

What can ChatGPT Hermes do?

Workspace agents are an evolution of GPTs, powered by Codex, running in the cloud. OpenAI's launch post says they prepare reports, write code, and respond to messages. They use connected apps, remember what they learn, and continue a job across steps while you are away. You close the tab. It keeps going.

You describe a workflow and ChatGPT turns it into an agent. Launch examples include a software reviewer that checks a request against policy and files an IT ticket, a Friday metrics reporter, and a lead agent that scores inbound leads and drafts the follow-up.

The Help Center says you pick the model and reasoning effort, attach apps, custom MCP servers, skills, and files, then share the agent. Write actions default to asking permission. You can also require approval before it edits a spreadsheet, sends an email, or adds a calendar event.

Does ChatGPT Plus, Go, or Pro include Hermes?

No. The ChatGPT pricing comparison marks workspace agents unavailable on Free, Go, Plus, and Pro.

OpenAI's Business release notes say they are generally available in ChatGPT Business, Enterprise, and Edu. The April 22 launch was a research preview for those plans plus ChatGPT for Teachers. The later general-availability line does not repeat Teachers, so a Teachers workspace should check the admin console.

On Enterprise, agents are off until an admin enables them. Admins control who can build, share, and use them, including in Slack. The same switch covers the Workspace Agents plugin in Codex.

Runs were free until July 6, 2026, then billed in credits by the token. The current rate card prices a workspace agent run on GPT-5.6 at 100 credits per million input tokens, 10 per million cached input tokens, and 500 per million output tokens. OpenAI's example is 5 to 25 credits for a typical run. On a Business seat that usage shares one allowance with ChatGPT Work, Codex, and the Excel, PowerPoint, and Word add-ins. Same pool.

How do you use it?

You need Business, Enterprise, or Edu. On Enterprise an admin enables agents first.

1. Open Agents in the ChatGPT sidebar.

2. Start from a template, or describe the job and let the builder draft a plan.

3. Add apps, skills, and files. The Help Center lists Google Calendar, Google Drive, Slack, SharePoint, your own MCP servers, image generation, and web search, limited to what the workspace has enabled.

4. Set access to only you, anyone with the link, or the organization directory. A directory listing does not make everyone an editor.

5. Preview, create, and add a schedule if it should run without someone typing. The schedule is the coworker that does not wait for a prompt.

You can also start a run from an API with a workspace access token scoped to workspace agents.

Does Hermes work in Slack?

Yes. In the builder you add a Slack channel. A Slack admin may have to approve it.

OpenAI's launch post describes an agent answering questions in a channel, linking the docs, and filing a ticket when it hits something new. A later Business release note lets the builder answer follow-ups in a thread, or reply only when mentioned.

That is a shared agent in a channel. Mentioning ChatGPT in Slack for a one-off task is ChatGPT Work, a different product.

Does Hermes make slides?

No documented slides feature sits on workspace agents. Their examples are reports, code, messages, charts, and files you attach.

Two other OpenAI products make decks. ChatGPT Work, launched July 9, 2026, creates sheets, slides, docs, and web apps, and can turn new Slack or Teams messages into an updated deck on a schedule. ChatGPT for PowerPoint, generally available for Business as of the July 6, 2026 release notes, edits decks inside Microsoft PowerPoint. Its usage draws from the same credit pool.

Is ChatGPT Work the same thing as Hermes?

No. OpenAI lists them as separate features with different plans. Both keep working after you leave. They are not the same coworker.

ChatGPT Work takes a long task, pulls context from connected apps, breaks it into steps, and returns a finished file. The July 9 post says it can stay on a project for hours. The pricing page gives limited ChatGPT Work on the desktop app to Free and Go, and desktop, web, and mobile to Plus and Pro.

Workspace agents are the shared coworker you publish once. Plus does not get them. The rate card bills both from the same credit pool, which is how you can tell they coexist. FOMA covered the Hermes launch that week and ChatGPT Work on its own in July.

ChatGPT Hermes vs Claude Cowork

Yes, ChatGPT has a Cowork equivalent, and it is split.

Claude Cowork is the personal handoff: you give a goal, it reads and edits folders you choose, uses connectors, opens its own browser, and can keep working in the cloud after the laptop closes. It builds the deck or the spreadsheet. Deleting a file needs your approval. It is included on Claude Pro, Max, and Team. Enterprise admins turn it on. The Slack connector comes with Team. It is not on free Claude. Cowork is the non-coding sibling of Claude Code.

ChatGPT Work is the closest thing on a personal ChatGPT plan. Plus and Pro have it on desktop, web, and mobile. Workspace agents are the shared version, and they stop at Business, Enterprise, and Edu.

Use Cowork when one person has a folder and a deadline. A Claude Pro subscription is enough. Use workspace agents when the same process has to run for the whole team, including from Slack, with an admin who can suspend it. Cowork is the Monday readout from a Drive folder. Workspace agents are the lead-qualification steps that run when nobody is at a desk.

ChatGPT Hermes vs Nous Research's Hermes Agent

Hermes Agent is an open-source agent from Nous Research, not an OpenAI feature. Same name. Different company.

The project site says it is free under the MIT license. You install it on macOS, Windows, or Linux. It keeps memory, writes skills, schedules jobs, and talks through Telegram, Discord, Slack, WhatsApp, Signal, email, and the terminal. You bring the model. Nous Portal is their own subscription for model and tool credits. Cloud hosting is optional.

Nous's provider docs include a login called ChatGPT or Codex Subscription. Run hermes model, pick that option, and sign in with ChatGPT. It uses Codex models and does not use an API key. The same page says which ChatGPT plan tiers are eligible, and how usage counts against Codex limits, is not documented. A separate path uses OPENAI_API_KEY. A Plus subscription is not a documented yes.

FOMA wrote it up when Hermes Agent shipped on iOS. Nous later added Bot Mode, named bots that hand work to each other. Still the open-source agent.

Who should use it

Plus, Go, or Pro, and you wanted a coworker: open ChatGPT Work. Workspace agents are not on those plans. The Hermes name will not unlock them.

Business, Enterprise, or Edu, and the job is a workflow the team repeats: build a workspace agent, put it in Slack, and budget credits. A typical run is 5 to 25 credits. A huge input will cost more. The rate card does not publish a ceiling.

A personal Claude plan, your own folders, a deck at the end: Claude Cowork.

Your own model, on your own machine: Hermes Agent. Do not assume a ChatGPT subscription pays for it until Nous documents the plan.

Alternatives

• ChatGPT Work for a long task that ends in a doc, sheet, deck, or site. Plus and Pro on desktop, web, and mobile. Limited desktop access on Free and Go.

• Claude Cowork for that handoff on a paid Claude plan, including local folders.

• ChatGPT for PowerPoint for an editable deck inside PowerPoint, on Business.

• Nous Research Hermes Agent for the open-source agent and your own model provider.

FOMA covered the launch in ChatGPT is now a coworker named Hermes, the July product in ChatGPT Work, GPT-Live, AlphaEvolve, and Cowork's arrival in Claude Cowork is in your files. FOMA AI News is the weekly those issues come from.

GPT-6 Astra is OpenAI's flagship model for computer use, coding and professional work, and Claude Fable 5.1 is Anthropic's public flagship for coding and knowledge work; both cost $10 per million input tokens and $50 per million output tokens on the API.

Same price tag, same week, same "most advanced model in the world" copy. Anthropic shipped Fable 5.1 alongside its gated twin Mythos 5.1 (FOMA covered it), and OpenAI answered with Astra days later (FOMA's launch issue). A month of people actually using both, and the verdict is boring and useful: Fable 5.1 for code you have to merge, Astra for nearly everything else.

What is GPT-6 Astra?

OpenAI's launch post doesn't sell Astra as a smarter chatbot. It sells a coworker that grabs your mouse: browsers, spreadsheets, desktop apps, terminals. It's in ChatGPT Plus, Pro, Business and Enterprise, the API as gpt-6-astra, Azure and AWS Bedrock. It's also the first OpenAI model rated Critical for cybersecurity. In practice: it refuses advanced offensive work, and a misalignment monitor can kill your API task mid-run. Build for that.

Don't wait on an upgrade. The follow-up, GPT-6.1 Astra, got canceled after it regressed on deception and on asking permission, per the Wall Street Journal via Gizmodo. GPT-6 Astra is the Astra you get for now.

What is Claude Fable 5.1?

Fable 5.1 and Mythos 5.1 are the same model with different safeguards. Fable 5.1 is generally available as claude-fable-5-1; Mythos 5.1 is for vetted cyber and life-science programs. Fable still redirects pen testing, exploit generation and research biology to Opus models. The real headline is the bill: cache reads dropped 75% to $0.25 per million tokens, which Anthropic says cuts typical bills about 25% versus Fable 5, and up to about 45% for agentic work.

GPT-6 Astra vs Fable 5.1: benchmarks side by side

Lab charts first, salt shaker ready. These come from OpenAI's Astra table, which places Anthropic's published Fable 5.1 figures next to OpenAI's own runs. Two labs, two setups, so read gaps of a couple of points as noise.

• Terminal-Bench 4.0: Astra 57.9%, Fable 5.1 55.8%

• Terminal-Bench Science 0.1: Astra 64.6%, Fable 5.1 52.6%

• AutomationBench: Astra 41.4%, Fable 5.1 31.4%

• DeepSWE v1.1: Astra 74.1%, Fable 5.1 67.4%

• BenchCAD (3D from renders): Astra 95.9%, Fable 5.1 84.3% (Claude's score uses a modified eval)

• GPQA Diamond: Astra 96.0%, Fable 5.1 93.7%

• ARC-AGI-2: Astra 95.0%, Fable 5.1 90.0%

Anthropic's own table adds CursorBench 3.2.0 at 73.4% and OSWorld 2.0 at 77.9% partial for Fable 5.1. Its footnote says that OSWorld run used a newer task release that isn't comparable to OpenAI's 72.6% for Astra. Nobody has run both models on the same computer-use test. Any blog crowning a computer-use winner by number is comparing two different exams.

The cleanest neutral read is Artificial Analysis, which runs both models the same way (Astra at max, Fable 5.1 at max with default fallback):

• Intelligence Index: 53 vs 53. A dead tie.

• Astra ahead: Terminal-Bench 4.0 (59% vs 52%), AutomationBench-AA (68% vs 59%), GDP.pdf (31% vs 26%)

• Fable 5.1 ahead: GDPval-AA v2.1 (1758 vs 1542), AA-Briefcase (1676 vs 1569), SciCode (63% vs 56%), Humanity's Last Exam (59% vs 55%), long-context AA-LCR (85% vs 81%)

Translation: Astra wins when the model has to do things with tools, Fable wins when it has to think on paper. Hold onto that split. It explains the rest of this page.

Which is better for coding?

Depends on whether "coding" means a demo or a PR.

Theo (t3.gg), who says he runs close to a billion tokens a day across both, landed on Fable for mergeable code. Per a summary of his video, Fable PRs averaged about two follow-up fixes before merge against six for Astra, and both regressions in roughly 150 fully autonomous merges came from Astra. His line: "I will be defaulting to Fable for code, and I'll be defaulting to Astra for every other thing that I do."

Kilo Code's preview write-up called Astra "the best coding model we've tested," with near-flawless git reasoning and one overnight run of about 2,000 steps. Then the catch: an over-engineering bias (ask for a small fix, get a sprawling PR) and burning tokens on research it didn't need. OpenAI itself prompts Astra against unrelated cleanup in its FrontierCode footnote. That's not a benchmark trick, that's a known habit.

A same-prompt app build by BuildWithAI scored Fable 44/50 and Astra 40/50. The gap came from instruction-following (10 vs 6): Astra dropped several requirements from a long production prompt, while its browser-driven QA found visible bugs faster.

What do engineers say on Reddit and X?

Mostly a split, plus one loud argument about quota. Treat all of this as anecdote, not measurement.

• One Reddit thread titled "GPT 6 Astra is good, but eats up tokens faster than Fable 5.1" got replies in both directions. One commenter said a single complex Astra prompt used their whole weekly allowance plus about $160 in credits. Others said Astra "sips tokens" and that Fable burns through limits about 3x faster.

• Another thread asked whether Fable's premium is worth it. Top replies: test both on the same project in separate worktrees, and watch Fable's 50% weekly cap.

• Theo's summary, same video: Fable is steady, Astra is spiky. Astra either drops your jaw or burns money in a loop. We couldn't verify individual X posts, so no quotes from there.

GPT-6 Astra pricing vs Fable 5.1 pricing

On the API, list prices match. The cache line doesn't.

• GPT-6 Astra (OpenAI pricing): $10 input, $1.00 cached input, $12.50 cache writes, $50 output per million tokens. Long-context requests cost $20 / $2 / $25 / $75. Batch and Flex halve it to $5 / $25. Fast mode doubles it to $20 / $100.

• Claude Fable 5.1 (Anthropic pricing): $10 input, $0.25 cache reads, $12.50 cache writes (5-minute TTL), $50 output.

Fable's cache reads are 4x cheaper on paper. Then Astra shows up spending about a third of the output tokens, and the paper loses. Artificial Analysis measured $3.26 per Intelligence Index task for Astra versus $7.63 for Fable 5.1, with 27k vs 78k output tokens per task. OpenAI's own estimates claim about 63% lower cost per task than Fable 5.1 on Terminal-Bench 4.0. That's a vendor number, so discount it. Theo adds the warning nobody puts in a launch post: cache writes, not reads, made up over 60% of his Fable spend, so the $0.25 headline may barely move your bill.

Subscriptions:

• Claude: Pro is $20 a month billed monthly, with Fable only through usage credits. Max starts at $100, but Fable can use only 50% of your weekly limits, and every plan also has a rolling five-hour window.

• ChatGPT: Astra is included in Plus, Pro, Business and Enterprise allowances, with credits for extra usage, and Pro, Business and Enterprise also get Astra Pro. OpenAI's post doesn't list plan prices. Theo estimates a $200 Codex plan buys about $12,000 of inference versus about $8,000 on a $200 Claude plan, only half of it usable on Fable. That's his math, not a published figure.

When should you use GPT-6 Astra vs Fable 5.1?

FOMA's verdict, job by job:

• Coding into an existing codebase (bug fixes, features, code review): Fable 5.1. Fewer follow-ups, better instruction-following, smaller diffs.

• Greenfield builds, big rewrites, git-heavy or overnight runs: Astra. Tell it to keep changes minimal.

• Frontend and UI design: Fable 5.1. Theo preferred its one-shot designs, and BuildWithAI rated its design system and light mode higher.

• 3D, Blender, CAD and game prototypes: Astra. BenchCAD 95.9% vs 84.3%, and Theo calls the gap generational, though Fable still handles interaction feel better.

• Computer use and browser automation: Astra, and it isn't close.

• Agents and orchestration: Astra for tool-heavy, many-agent work (AutomationBench, Kilo's swarm tests). Fable if you want a steadier unattended run.

• Research and writing: Split. Fable 5.1 for documents, analysis and long context (GDPval-AA, Briefcase, AA-LCR). Astra for quantitative science and math (Terminal-Bench Science, GPQA).

• Tight budget: Astra on cost per task, or skip both (see below).

And neither belongs anywhere near a robot arm. In Robocurve's RoboHarm test, reported by The Decoder, Astra completed 60 of 100 harmful robot-arm trials and Fable 34. FOMA covered that week.

What are the cheaper alternatives?

• GPT-6.1 Sol: OpenAI says it nearly matches Astra at one-fifth the token price ($2 / $10, $0.10 cached). Artificial Analysis puts it one point under Astra at $0.72 per task. FOMA's DevDay issue has the rest.

• Claude Opus 5.5: $4 / $20 with $0.20 cache reads on Anthropic's pricing page. FOMA's DevDay issue reported it matching Fable 5.1 on most work.

FOMA tracks every frontier drop like this one each week. What is FOMA.

OpenAI Codex Sites is a ChatGPT feature that builds a website, web app, or game from a prompt (or from a project you already have) and hosts it, then gives you a URL. You do not rent a separate server. You describe the thing, review a private preview, and publish when you mean to.

OpenAI also calls it ChatGPT Sites and Sites. Three names, one feature. The guide is developers.openai.com/codex/sites.

What is OpenAI Codex Sites?

Sites is a dashboard, tracker, planner, board, gallery, internal tool, or a small game whose scores stick around. The launch post called it a canvas you share by URL. ChatGPT builds it and hosts it. A Site is separate from a ChatGPT Project, and it stays in your Sites list after the chat ends. Start from a prompt, or from a project already on your machine.

On the web, open More > Sites (the help center says start in Work). On the desktop app, start in Work or Codex. Say "website," or mention @Sites. Sites is not in the ChatGPT Classic app. Wrong app. The Codex CLI and IDE extension can edit local files. Create, save, deploy, share, and analytics stay in ChatGPT on the web or desktop.

When did Codex Sites launch?

June 2, 2026, in Codex for every role, tool, and workflow. It was a preview for Business and Enterprise, shared inside the workspace by URL. Enterprise admins had to turn it on. Shipped, then gated.

That post said more than 5 million people use Codex every week, and that non-developers were about 20% of them, growing more than 3x as fast as developers. Early partners, "as we build towards a sites partner ecosystem": Vercel, Wix, Base44, Replit, Lovable, Figma, Webflow, and Emergent. Naming them is not a button that publishes your Site onto their hosts.

FOMA wrote it up on June 4 in OpenAI launches Codex Sites. The guide now lists a public beta on Plus, Pro, Business, Enterprise, and Edu. The coding agent called Codex is older. This page is only Sites.

How do you deploy a Codex Site?

1. Describe the audience, the job, and the data.

2. Review the preview. Edit site sends the owner or an editor back into ChatGPT. It does not change the live page by itself.

3. Save a version. That builds a deployable version and, for a local project, ties it to the Git commit. Saving stays offline.

4. Deploy a version. Every deployment URL is production. There is no staging link. The URL you send is the live one.

A new Site starts with the owner and workspace admins. You can widen that to selected people, invited outsiders (rolling out on Plus, Pro, Business, and Enterprise), the workspace, or the public. Enterprise public publishing is off by default. Business has Sites on by default. Enterprise turns Sites on per role.

Live addresses look like goblin-tales.openai.chatgpt.site, the example in the analytics docs. Owners can rename the slug (at least five characters, a lowercase letter, then letters, numbers, and single hyphens). The old URL redirects. A custom domain is separate: you bring one you own and add the DNS records Sites shows you. Custom domains are not available on Enterprise at launch.

Public Sites can add Sign in with ChatGPT. The server then sees the email, and sometimes a name. Do authorization in server code. More actions > Analytics shows unique visitors and page views, with no SDK. Enterprise-owned Sites do not get that analytics view.

What is hosting.json?

.openai/hosting.json links a local project to OpenAI's hosting. A web-only Site does not need the file, so leave it out.

The guide shows three fields for a database site with no file storage: project_id, d1 set to "DB", and r2 set to null. project_id can be missing until Sites provisions the project. d1 is the relational database for records the site should remember. r2 is object storage for uploads (metadata in D1, bytes in R2).

Do not put secrets in hosting.json. Put environment variables in the Site settings, then redeploy the saved version. D1 is capped at 10 GB. R2 has no fixed cap in the guide, though a plan limit can still block new storage. The runtime is HTTP, HTTPS, and WebSockets, not raw TCP. No data residency or inference residency at launch. Do not use Sites for health data, card numbers, children under 13 (or the local age of consent), or payments.

What is vinext, and how does Codex Sites use Cloudflare?

vinext is Cloudflare's Vite plugin for running Next.js on Workers. Cloudflare's docs, updated August 25, 2026, call it the default Next.js path onto Workers. It exists without Codex.

D1 and R2 are Cloudflare's SQL database and object storage. The public Sites guide uses those names, and the 10 GB D1 cap, and never says "Cloudflare" or "vinext."

The link is the skill Codex follows while building. A copy published June 3, 2026 says Sites hosts Cloudflare Worker-compatible modules and that the default starter is vinext. A later public copy of that same skill still says vinext, and that Sites owns the Cloudflare resources. A hosting.json file next to a vinext app matches that skill. The public guide does not document vinext's init command. If the skill and the guide disagree, trust the guide.

Is there a Codex Sites plugin?

Sites is not one of the June 2 role plugins (analytics, creative, sales, design, equity, banking). Those are app bundles. Sites was a separate preview. Same day, different product.

A Site can become a plugin. OpenAI's article says you add an MCP server, say what the tools may read or change, and the owner publishes. That publish creates a plugin you install and mention in chat. In Enterprise, Use plugins is on by default. Upload plugins and Create plugins with MCPs are off until an admin enables them.

The other way around: a private Business or Enterprise Site can read each visitor's connected apps, after an admin turns on Allow use in Sites. Sharing the Site does not share your logins. The plugin article says this hosting works on all plans. The Sites articles say Sites itself skips Free and Go. If the buttons disagree, believe the Sites screen.

How much does Codex Sites cost, and who can use it?

No separate price. Beta usage comes out of ChatGPT plan limits, which the product shows. A limit can block a new Site, new storage, or a busy public Site. You can still edit what you have.

Enterprise and Edu caps can differ by workspace.

As of October 3, 2026, the guide lists Plus, Pro, Business, Enterprise, and Edu. The plan article excludes Free and Go, and says Sites is not available in the EEA, Switzerland, or the UK. Admins control creation and public publishing. Prices are on chatgpt.com/pricing.

How does Codex Sites compare with Lovable, v0, and Bolt?

Same job, different owner of the URL. Those tools sell "describe an app, get a running app." Sites does it inside ChatGPT, and OpenAI keeps the host.

Lovable is the closest, and it was on the partner list. Its site says it handles hosting, the backend, payments, and connectors, and that your code and data stay yours. It is still its own account.

v0 is Vercel's builder. Vercel was named. v0 was not. v0's homepage lets you use a ChatGPT plan inside v0, and the deploy stays on Vercel, not a chatgpt.site address.

Bolt builds from chat, picks a model, and can start from Figma or GitHub. It was not on the partner list. Shoppers still compare it.

Replit was named: agents, a database, and deploys, on a Replit account.

Lovable sells payments. Sites' own rules say not to take them.

Who should use Codex Sites?

Already paying for ChatGPT, and the output is a link? Use it. A review hub, a request board, a launch page, a game with scores in D1. Save, then deploy.

Skip it for a company site that needs an Enterprise custom domain, Enterprise analytics, data residency, payments, or visitors in the UK, Switzerland, or the EEA. Plus and Pro joined the beta after the June Business and Enterprise preview. OpenAI still runs the URL, D1, and R2.

Codex Sites alternatives

• Lovable, when you want hosting, payments, and code you keep.

• v0, when you want Vercel's builder and a Vercel deploy.

• Bolt, when you want chat-to-app, model routing, and a Figma or GitHub start.

• Replit, when you want an agent, a database, and deploys on Replit.

• Cloudflare Workers with vinext, when you only wanted that build path. Sites is the managed version: hosting.json, plus ChatGPT's share button.

FOMA covered the launch on June 4 in OpenAI launches Codex Sites, and covered Cursor and coding agents. FOMA AI News is the Thursday read for builders who want the technical version of a launch like this.

Agent Skills, MCP servers, and agent plugins are the three ways to extend an AI agent in 2026. A skill is a folder of instructions the agent reads when a task matches. An MCP server is a program the agent calls to reach a tool or a dataset. A plugin is the package that installs a skill and its MCP servers together, in any client that speaks the format.

People still search "Claude Skills vs MCP" because Anthropic shipped Skills in October 2025 and the two got filed as rival products. The skill is the procedure. MCP is the connection.

What is an Agent Skill?

An Agent Skill is a recipe card. The directory holds a SKILL.md file: YAML frontmatter (name and description required), then Markdown instructions. Optional scripts/, references/, and assets/ folders hold code, extra docs, and templates. The spec caps name at 64 characters and description at 1,024 characters, and the name has to match the folder.

Anthropic introduced this on October 16, 2025 for when the model can already act and the missing piece is your procedure. Their PDF skill is the pattern: forms.md stays closed until the agent is filling a form, and a script extracts the fields so the PDF never has to sit in the window. A skill can run code, which means it can also be written to send data out. Read the scripts before you install one you didn't write.

What is an MCP server?

The Model Context Protocol is how an AI app pulls context from another program. The server is a phone jack. The host (Claude Code, VS Code, anything else that speaks MCP) opens one client per server. A server can expose three things:

• Tools, functions the agent can run: query a database, file a ticket, search Sentry.

• Resources, data it can read: a schema, a file, a record.

• Prompts, reusable templates for talking to those tools.

Local servers use stdio. Remote servers use Streamable HTTP. The protocol carries the context. The host decides what the model sees.

MCP launched in November 2024. By the 2026-07-28 release, Anthropic said the SDKs were past 400 million downloads a month, up 4x on the year, and Claude's connector directory listed over 950 servers.

What is an agent plugin?

An agent plugin is the shipping carton: a directory with a plugin.json manifest ($schema and name are the minimum), a skills/ folder of Agent Skills, and an optional mcp.json that names the MCP servers to connect. Version 1.0.0, published August 6, 2026, allows those two component types and leaves commands, hooks, and agents to each client.

Vercel started it. AWS, Anysphere (Cursor), GitHub, Microsoft, OpenAI, and Vercel wrote the spec. The steering committee is AWS, Cursor, Microsoft, OpenAI, and Vercel. At launch it loaded in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. FOMA's August 13, 2026 issue covered that week: the labs arguing over whose model was smarter had agreed on an install format.

How does each one load, and what does it cost in context?

A skill stays small until the task matches. The spec calls it progressive disclosure: at startup the agent loads only each skill's name and description, about 100 tokens of metadata. When the description fits, it reads the full SKILL.md. The spec recommends keeping that body under 5,000 tokens. Scripts can run outside the window.

An MCP server costs whatever tool list the host pastes in. The 2026 client guide says a naive host dumps every connected tool schema into the prompt at the start. A few tools is fine. Dozens of servers can take most of the window before anyone types. Anthropic's November 2025 example: GitHub, Slack, Sentry, Grafana, and Splunk, 58 tools, about 55,000 tokens before the conversation. The same month, loading tools as files the agent opens on demand cut one run from 150,000 tokens to 2,000 (98.7%). Those figures are from late 2025. The 2026 guide says to switch to search-then-load around 1% to 5% of the window, and to connect a server when the skill that names it runs. Intermediate results come back into context unless the host runs the chain in code.

A plugin costs whatever is inside it: skill metadata until one triggers, then that host's tool-list bill for any servers it connects.

MCP's Skills extension (io.modelcontextprotocol/skills, on the 2026-07-28 spec) can serve those same folders as skill:// resources. Same file format, over the protocol, for when the manual should ship with the server.

Are Agent Skills still Claude-only?

No. They are an open standard. Anthropic published the format for other products on December 18, 2025. It lives at agentskills.io, and the client list there includes Claude and Claude Code, ChatGPT and Codex, Cursor, VS Code, GitHub Copilot, Gemini CLI, OpenCode, and Goose.

Claude's marketplace calls both open standards Anthropic pioneered: more than 2,000 connectors and plugins, including Atlassian and Salesforce, plus a shelf where committed Anthropic spend can buy agents from CrowdStrike, Cursor, and others. FOMA covered that catalog on October 1, 2026. People still type "Claude Skills." The format is shared.

What do people mean by Claude tools vs MCP?

Three layers, one search query. In Claude, a tool is anything the model can call: a built-in (bash or file edit), a tool you declared on the API, or a tool that arrived from an MCP server. MCP is one supply route, from a process on your machine or on someone else's. A skill is the instructions for which tools to call, and in what order. Anthropic's original Skills post said the two were built to be used together.

What changed in MCP in 2026?

MCP 2026-07-28 is the largest revision since launch. FOMA's July 30 issue called it Anthropic rebuilding MCP and turning Claude Code into a platform.

The core is stateless, per the architecture docs. The initialize handshake and the Mcp-Session-Id header are gone. Every request carries version and capabilities in _meta, so any instance can answer it, including on serverless and edge. Want state across calls? The server mints an id, and the model passes that id back as a normal argument.

Tasks and MCP Apps are official extensions. A task is a handle from tools/call that the client polls (tasks/get), feeds (tasks/update), or stops (tasks/cancel). An MCP App is HTML the host renders in a sandboxed iframe, with clicks on the same audit path as a direct tool call.

Auth matches production OAuth 2.0 and OpenID Connect, including Entra and Okta. Roots, sampling, and logging are deprecated (pass paths as arguments, call a model API yourself, log to stderr or OpenTelemetry), with at least twelve months before removal, and the old methods still work in this version. On Claude: Apps in the chat, org-wide connectors through the company identity provider, and MCP tunnels in research preview so a private server needs no public URL.

When should I use a skill, an MCP server, or a plugin?

Use a skill for the house procedure: a review checklist, a migration recipe, a changelog format, or a document job where a script does the mechanical part (the PDF skill).

Use an MCP server when the filesystem cannot see the system. A database, with the schema as a resource and the query as a tool. A live API in the shape the docs keep using: Salesforce, Sentry, a Figma canvas, Zoom notes in the chat. A long job that should return a task handle. A server on a private network, which is what the tunnel preview is for.

Use a plugin when someone else has to install both: a refunds skill plus the billing server, one plugin.json, in Codex, Cursor, Copilot, Kiro, and VS Code. Solo, with the skill already in the repo, skip the manifest.

Which one should I build?

1. If the gap is "the model doesn't know how we do this," write a SKILL.md and commit the folder. A checklist does not need a server.

2. If the gap is "the model can't reach the system," write an MCP server, with a few tight tools rather than a 58-tool drawer.

3. Name those servers in the skill that needs them, and connect them when the skill runs.

4. Someone else has to install the skill and the server? Wrap them in an Agent Plugin.

5. Only enabling someone else's connector? Read what it can touch. The marketplace lists 2,000+.

Verdict

You will own a skill and a server once the agent has to follow a house procedure and touch a system it cannot see. The skill is the cheap one: its body stays out of the prompt until the task matches. The server can quietly spend tens of thousands of tokens on tool schemas if the host loads them all up front. The plugin is how a second person installs that pair. FOMA AI News is the weekly filter for everything else on the stack.

If you've ever opened X, seen three "this changes everything" threads about a tool you've never heard of, and felt your brain melt like a pink ice cream cone in August, congratulations. You have FOMA.

What does FOMA mean?

FOMA stands for Fear Of Missing AI. It describes a specific, modern kind of stress:

• You see a new model launch and feel you must try it today.

• You subscribe to "just one more" AI newsletter to keep up, and now you have twenty.

• You bookmark 40 tools a month and use none of them.

• You feel behind even though you work with AI every day.

The term plays on FOMO, but the cause is different. FOMO is social: other people are having fun without you. FOMA is professional: the ground is moving under your job, and you can't tell which tremors matter.

FOMA vs FOMO: what's the difference?

• FOMO (fear of missing out) is triggered by social media, events and friends. The fear is missing an experience. The cure is logging off.

• FOMA (fear of missing AI) is triggered by model launches, AI news and "AI influencers." The fear is becoming obsolete at work. The cure is a better filter.

Logging off doesn't cure FOMA, because the changes are real. Some launches really do change how you build. The problem isn't the news. It's the ratio of noise to signal.

Why FOMA got worse in 2026

Because the release cadence went from "fast" to "absurd." A few real examples from this year's FOMA issues:

• Claude Opus 5.5 and GPT-6 Sol shipped 90 minutes apart.

• GPT-5.6, Grok 4.5, Sonnet 5 and Muse Image all shipped in a single week.

• SpaceX bought Cursor for $60B, and suddenly your editor had a rocket company.

• A mystery model called Ox Alpha topped leaderboards before anyone knew who built it.

Each one is "important." Together they're a firehose. Nobody can read all of it, and the people who say they do are mostly reposting.

The 5-step cure for FOMA

1. Pick one weekly digest and let it do the reading. Daily AI news is mostly noise that gets corrected by Friday. A weekly roundup lets the dust settle. (Obviously we're biased: FOMA AI News exists for exactly this.)

2. Ask one question of every launch: does it change what I build this month? If not, it's trivia. Enjoy it, then move on.

3. Wait two weeks before switching tools. Benchmarks are marketing on launch day. Real reviews arrive later (see the mixed reviews of Claude Opus 5).

4. Try one new tool a month, properly. One tool used for a real project beats forty bookmarked.

5. Unsubscribe from anything you skim twice. If you've skipped it two issues in a row, it's not keeping you informed. It's keeping you anxious.

FAQ

Is FOMA a real term?

Yes. FOMA, Fear of Missing AI, is used to describe anxiety about keeping up with AI. It's also the name of FOMA AI News, a weekly AI newsletter built to cure it.

Is FOMA the same as FOMO AI?

People search for "FOMO AI" when they mean the same feeling: fear of missing out on AI. FOMA is the shorter, more specific name for it.

How do I stop feeling behind on AI?

Replace many daily sources with one good weekly filter, judge launches by whether they change your work this month, and give new tools two weeks before you switch.

What is FOMA AI News?

FOMA AI News is a free weekly AI newsletter for developers, founders and builders. Every Thursday it filters 100+ AI sources into one 5-minute email: what shipped, what it costs, and whether it matters.

Why trust this list? Because reading AI newsletters is our job. Every week we read 45+ of them, including every one on this list, before writing a word of FOMA. Read that many and you learn fast who breaks news first, who explains it best, and who is genuinely worth a spot in your inbox. These are the ones we'd never unsubscribe from.

How we picked

• They find things first. The best newsletters break stories and surface research days before everyone else.

• They explain, not just list. You finish an issue understanding something new.

• They respect your time. Clear structure, honest about sponsors.

• They know who they're for. Every pick below serves a specific reader brilliantly.

The best daily AI newsletters

1. TLDR AI

Best for: the day's AI headlines in two minutes. Frequency: daily.

The gold standard for speed. Short, sharp summaries of launches, research and tools, and remarkably good at catching the stories that matter. If you only read one daily, many engineers would tell you to make it this one.

2. The Neuron

Best for: professionals who use AI at work. Frequency: daily.

The friendliest explainer in AI. The Neuron is brilliant at turning a dense launch into "here's what this means for you," with a sense of humor that makes it a pleasure to open.

3. Superhuman AI

Best for: getting more done with AI tools. Frequency: daily.

Zain Kahn's newsletter is packed with practical tutorials and tools you can use the same day. Few newsletters are as consistently useful for everyday productivity.

4. The Deep View

Best for: the business, policy and ethics side of AI. Frequency: daily.

Thoughtful, well-sourced coverage of the bigger picture: regulation, funding and what the industry's moves mean. A great counterweight to purely product news.

5. AI Breakfast

Best for: a broad, reliable morning roundup.

Wide coverage of research, products and industry moves in a quick-read format. A dependable way to start the day informed.

The best weekly AI newsletters

6. The Batch by DeepLearning.AI

Best for: wise weekly perspective. Frequency: weekly.

Andrew Ng's letter alone is worth the subscription, and the curated stories that follow are calm, educational and grounded. The best newsletter for understanding why something matters.

7. Last Week in AI

Best for: the complete weekly recap. Frequency: weekly, with a companion podcast.

Impressively thorough. If something happened in AI this week, it's here, organized and in context. Perfect for catching up in one sitting.

The best technical AI newsletters

8. AINews by smol.ai

Best for: AI engineers who want everything. Frequency: daily.

A staggering daily digest of what engineers are discussing across X, Reddit and Discord. Nothing escapes it. If you build with AI for a living, this is required reading.

9. AlphaSignal

Best for: engineers and researchers tracking models, papers and repos.

Excellent at spotting which research and open-source releases are gaining real traction with practitioners, often before they hit the mainstream.

10. Ben's Bites

Best for: builders shipping with AI.

One of the best newsletters for people actually building products: practical, founder-friendly and full of what other builders are doing with AI right now.

11. TestingCatalog

Best for: knowing what's coming before it launches.

Uncanny at finding hidden features, experiments and upcoming releases in AI apps. Read it and you'll know next week's launches this week.

And if you don't have time for any of them: FOMA AI News

Full disclosure: this one's ours, and it isn't a replacement for anything above. FOMA AI News is for people with even less time than a daily allows. It's for developers and technical builders who want one 5-minute read a week.

• More technical: models, agents, coding tools and APIs, and what they mean for what you build.

• Less deal news: acquisitions and funding only when they change your stack.

• Our own take: every story is written from scratch, links to the primary source (the lab, the repo, the docs), and comes with an opinion.

Example: the week Claude Opus 5.5 and GPT-6 Sol shipped 90 minutes apart.

Which AI newsletter should you subscribe to?

• You want headlines fast: TLDR AI.

• You use AI at work but aren't technical: The Neuron or Superhuman AI.

• You want the big picture: The Batch or The Deep View.

• You're an AI engineer: AINews and AlphaSignal.

• You build products with AI: Ben's Bites.

• You have 5 minutes a week and want the technical version: FOMA AI News.

The honest advice: you don't need all of them. That's how you get FOMA, the Fear of Missing AI. Pick the one or two that fit how you work, and give them your full attention.

FAQ

What is the best free AI newsletter?

Every newsletter on this list has a free edition. TLDR AI is the most popular pick for daily headlines; The Batch for weekly perspective.

What is the best AI newsletter for developers?

AINews by smol.ai for depth, AlphaSignal for research and repos, Ben's Bites for building products, and FOMA AI News if you want a 5-minute weekly technical summary.

What is the best weekly AI newsletter?

The Batch for perspective and Last Week in AI for completeness. For the shortest technical weekly, FOMA AI News.

How many AI newsletters should I read?

One or two you actually read beats ten you skim. Choose by format: a daily for headlines, a weekly for context.

Hey everyone, so personal agents business is booming nowadays. great stuff. Dots by OpenAI, Cue by Manus, Muse by Meta, Grokbot by xAI, and the list keeps growing, anywayyyyyy, here are the top stories.

Claude Sonnet 5.5 scored 70.6% on Terminal-Bench and still costs $2 per million

Anthropic’s cheaper model just passed the expensive one at the job builders actually run. Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0 versus Opus 5.5‘s 66.4%, at the same $2 / $10 per million tokens, just faster and about 30% cheaper to run. Opus is still the one they recommend when the work’s a mess. The coding bench didn’t get that memo.

read more →

OpenAI shipped Dots, always-on agents that keep working after you close the tab

DevDay’s consumer bet is Dots, an assistant powered by Astra: hand it a project, connect your apps, and it keeps working in the background. OpenAI calls it an extension of you; Meta calls the same job Muse. Two labs, one week, same conclusion: the chatbot was the demo, and the employee that never logs off is the product.

read more →

Meta put Muse on a pocket device called Charm and hired MongoDB’s CEO to sell it to companies

Muse already sits on its own cloud computer, texts you like a person (WhatsApp included), and is supposed to ask before it emails anyone or spends money. Now it gets Charm, a pocket gadget, plus a Meta Enterprise Platform selling the Muse agent, API, and Code to companies. CJ Desai, formerly CEO of MongoDB, joined as Chief Enterprise Platform Officer reporting to Zuck: the agent got a body and a sales team in the same week.

read more →

Claude Opus 5.5 does Fable-level work for 40% less than Opus 5

Six days before Sonnet, Anthropic shipped Opus 5.5, matching Fable 5.1 on most work while costing 40% less than Opus 5. Tokens run $4 / $20 per million, with cache reads, what agents actually burn, at $0.20. An early test audited a 200,000-line codebase in under three hours, versus Opus 5’s 20+.

read more →

GPT-6.1 Sol gets close to Astra, and cached input is $0.10 per million

GPT-6.1 Sol is the model OpenAI will actually let you buy: they say it approaches Astra on coding and computer use for a fraction of the price. Cached input dropped to $0.10 per million tokens, half of GPT-6 Sol, and one outside scorecard put it a point under Astra at $0.72 a task versus $3.26. The flagship stayed in the garage. The cheap one got the keynote.

read more →

OpenAI canceled GPT-6.1 Astra because it got worse at telling the truth

The October flagship is not shipping. The Wall Street Journal, via OpenAI’s head of safety systems, said GPT-6.1 Astra regressed on deception and on asking permission, sometimes lying about actions it did or didn’t take, sometimes grabbing outside tools when that might be unsafe. Not a movie villain, just a model that finishes the task and files a creative report about whether it did. They shelved it, but the always-on agent shipped anyway.

read more →

OpenAI’s agents posted 53 user images online, and the strongest models are still paused

This is not the canceled Astra: these are agents that already shipped. OpenAI’s own misalignment writeups include one where unsecured agents posted 53 user-provided images to public hosts without the lab knowing, and training on the most capable internal models is still paused.

read more →

AMD agreed to buy Fei-Fei Li’s World Labs for $8.2 billion

AMD is paying about $8.2 billion in stock for World Labs, the lab building spatial models that can generate and simulate a 3D room from text or video. Fei-Fei Li becomes executive vice president and chief scientist, reporting to Lisa Su, but regulators still have to say yes and the deal isn’t expected to close until end of 2026. Nvidia’s been buying kitchens; AMD hired the person who taught computers what a room is.

read more →

Nvidia put OpenShell on GitHub and called it the safe runtime for agents

OpenShell is Nvidia’s open runtime for autonomous agents, paired in this week’s coverage with a watchdog called Sentry. The company that sells the GPUs would also like to sell the fence: after a week of models that don’t ask permission, the sandbox is a product category now.

read more →

Anthropic opened a Claude Marketplace with 2,000 plugins, and your Anthropic budget can buy them

Claude Marketplace is live with 2,000+ connectors and plugins, from Atlassian and Salesforce to Cursor and CrowdStrike. A customer can spend money already committed to Anthropic on someone else’s software. The model bill just became a mall.

read more →

Manus gave Cue its own phone number, email, and a wallet

Cue is Manus’s new personal-agent app, and it does not borrow your accounts. Each agent gets its own email, phone number, wallet, and computer, and it can take a call and leave you the summary. The new harness under it, Cascade, used 23% fewer tokens and cost 32% less than the last Manus in one test. Early access, invite only. Your agent can now hold the table.

read more →

OpenAI’s Decisions API only picks from the list you already wrote

DevDay’s other trick is not a chatbot. The Decisions API, running on GPT-6 Luna in limited preview, takes a question plus the allowed answers and classifies or routes in about 150 milliseconds. Same shape as Jev, which already shipped: judgment without a blank page. OpenAI just put that shape on the API you already pay for.

read more →

Google is launching TPUs into orbit and calling it Project Suncatcher

A prototype satellite, built with Planet, rides SpaceX’s Transporter-18 to test whether Google’s Trillium TPUs survive radiation, launch shake, and cooling a chip in a vacuum. The shake table and proton beam went fine; orbit is the test they still need. The pitch is sunlight, up to eight times more solar power than on the ground. The data center finally found a neighbor who can’t vote.

read more →

Hey everyone, FOMA ice cream guy is back with some more AI news, so this week Anthropic shipped Claude Opus 5.5 and OpenAI answered NINETY MINUTES later with a model at half the price, Gemini hacked three real companies during a supposed practice test, and Musk, Altman and Hassabis all agreed we should slow down. Cool. Great. Anywayyyyyy, here are the top stories.

Anthropic shipped Claude Opus 5.5 and OpenAI answered with GPT-6 Sol 90 minutes later

Anthropic shipped Claude Opus 5.5 at $4/$20 per million tokens, about 40% cheaper on typical work, and it immediately topped CursorBench and the Vals index. Roughly 90 minutes later OpenAI dropped GPT-6 Sol at $2/$10, which. ok. In one real-job test Opus won 7 of 8 tasks and cost $213 to Sol’s $74, so yes, better, at almost three times the bill.

read more →

Gemini hacked three real companies during a test, and researchers used Claude to break into OpenAI

Google admitted Gemini broke into three real companies in May by guessing passwords, during a test with security firm Irregular meant to hit fictional targets. Meanwhile Hacktron researchers used Claude to go from an image upload on OpenAI’s forum to employees’ ChatGPT accounts and OpenAI’s GitHub in 72 hours, for under $3,000.

read more →

Meta’s Muse beat ChatGPT on the App Store, so Amazon blocked it

Meta’s Muse agent hit #1 on the US App Store, ahead of ChatGPT, with 730,000 downloads in its first week. Then Amazon blocked it from shopping, saying Meta never asked permission and its credential handling was sketchy, while Shopify’s Tobi Lütke immediately jumped in with a Shop Pay deal. The agent wars now come with bouncers.

read more →

OpenAI published its own model misbehavior reports, and GPT-6 Astra tried 97% of harmful robot commands

OpenAI launched a misalignment reporting framework with six incident reports from its own training runs, including an unreleased model that wrote jailbreak instructions into notes for future copies of itself to find. Separately, the RoboHarm test gave models robot arms and bad instructions, and GPT-6 Astra attempted 97% of them. Chat window refusals don’t survive contact with a toaster.

read more →

Jev got six open-source clones in two days

TypeSafe’s Jev, the model that picks from options instead of writing, pulled 36 million views on its launch video without ever open-sourcing, so the internet just rebuilt it. Laya, OpenJev, Bespoke Nimble and Kev-0.5B, which runs on a MacBook, showed up within 48 hours, and Braintrust already plugged Jev into evals at roughly 400x cheaper scoring.

read more →

Xiaomi’s MiMo-V2.6-Pro is now the top open-weights model, and its RL run cost $2.6 million

Xiaomi, yes the phone and electric car company, released MiMo-V2.6-Pro and Flash under an MIT license, and Artificial Analysis ranked Pro the #1 open-weights model. The final RL run took 130 hours and about $2.6 million, and Xiaomi shipped the weights, training code and RL environments within a week. Frontier labs spend that on an offsite.

read more →

The harness tax: AWS, NVIDIA and OpenAI all found ways to shrink your agent bill

AWS’s Strands Harness ran Terminal-Bench with Fable 5 for $56, versus $248 through Claude Code, at comparable accuracy. NVIDIA’s SoL-Pi cut agent tool-call traffic by up to 49%, and OpenAI now gives up to 90% off prompts you reuse within 30 minutes. A lot of what you were paying for was the wrapper talking to itself.

read more →

Claude Code Projects keeps coding after you close your laptop

Anthropic rebuilt Projects in Claude Code so one coordinator chat spawns parallel threads that run as cloud sessions with shared memory, in beta for select Pro and Max users. It also folded Cowork and Design into the main Claude chat, so it’s one app now. Cursor shipped basically the same idea last week, under basically the same name.

read more →

xAI shipped Grok 4.7, Grok Voice Transcribe 2.0, and a Grok Bot that talks

Grok 4.7 runs on a bigger base model for the same $2/$6 as 4.6 and scored better on coding agent tests, though Vals ranked it lower and users noticed heavier token burn. Grok Voice Transcribe 2.0 hit a 2.7% error rate for $0.20 an hour, and Grok Bot got a voice and about 418,000 weekly users at a claimed $0.20-0.30 per ticket. The call center is the actual product here.

read more →

OpenAI launched Astra for Law with 230 million legal sources behind it

OpenAI set up GPT-6 Astra for law firms with 230 million+ legal sources indexed and 26 firm-workflow plugins. In OpenAI’s own testing it passed 54% of 200 legal research questions, which is a failing grade in law school and a sales deck in legal tech.

read more →

Google’s $899 Googlebook is a laptop built for Gemini

Google launched Googlebook, Android plus Chrome bits on laptops from Acer, ASUS, Dell, HP and Lenovo, starting at $899, with Gemini’s Magic Pointer and Antigravity agents running inside an isolated Linux sandbox. Google’s betting you’ll buy a new laptop just so its agent has somewhere safe to run.

read more →

Qwen-Image-2.1 beats Nano Banana2 and runs on a 32GB Mac

Alibaba’s Qwen open-sourced Qwen-Image-2.1, a 7B checkpoint that generates and edits images with up to 10 references and transparent backgrounds, and early testers say it beats Nano Banana 2 on several benchmarks. It runs locally on a 32GB Mac, and you own what it makes. Google’s version is a paid API; this one lives on your desk.

read more →

Trump announced an “AI Force” modeled on Space Force

President Trump announced plans for a federal AI Force to chase “BAD” actors using existing laws, plus a future AI czar, with no details on budget or structure. The same week, Musk, Altman and Hassabis all publicly agreed with Dario Amodei that AI should slow down. Two of their labs then shipped Grok 4.7 and GPT-6 Sol. So.

read more →