Hey everyone, FOMA ice cream guy is back with some more AI news, so this week DeepSeek put a coding agent to work, Mistral made Le Chonk downloadable, and Google put Gemini 4 Carbon behind a guest list. Anywayyyyyy, here are the top stories.
DeepSeek opened a coding agent that comes back with pull requests
DeepSeek’s dsh-superagent watches a GitHub repository, takes requests, and gives jobs to separate worker sessions that can return pull requests. A panel flags work needing a human. The team calls it research code, not an official DeepSeek product, and says some of the grander designs aren’t implemented yet.
Google is shutting down Jules while Antigravity gets a plugin marketplace
Google is winding down its coding agent Jules: finish active sessions by November 11, then it becomes read-only until it closes on December 31. Google says the lessons are going into products including Antigravity, which also launched a plugin marketplace for its app and CLI. Jules users can export their settings and session history before the shutdown.
Mistral released Large 4’s weights, so you can stop renting Le Chonk
Mistral released improved Large 4 weights after last month’s preview, letting teams host and customize the model themselves under a modified MIT license. Mistral says its coding agent matched GLM-5.3 on one evaluation using 47% fewer tokens. The catch: businesses selling model access need a separate commercial license.
Canva Brush edits the whole design, and you can still move the pieces
Canva’s new Brush model takes requests to change an existing design, from a line of text to an entire presentation, while keeping the elements editable. It follows last week’s Canva Craft image generator, but works on the design around the image. Supported edits are available to Canva Pro subscribers, with wider access planned.
Google’s Gemini 4 Carbon is here, if you’re on the list
Google announced Gemini 4 Carbon, its strongest model yet, with gains in coding, finding security holes, and fixing them. Before you go looking for the dropdown: access starts with selected cyber defenders and expands through Google’s Fairwind program. Mistral is handing out weights while Google is checking names at the door.
MIT used GPT-6 Astra to fix another model’s reasoning loops
MIT researchers gave GPT-6 Astra Python access to a small Qwen model’s weights, and its edits lifted the model’s math score from 7.9% to 48.1%. Cool. It also got practice problems and repeated test runs; this wasn’t a psychic reading of a spreadsheet. It didn’t meaningfully improve the Gemma model they tested.
A malicious Claude Code extension slipped into Cursor’s marketplace
Cursor says tampered versions 2.1.312 and 2.1.313 of the Claude Code extension were available in its marketplace for about 11 hours on October 29. They attempted to plant an SSH backdoor; Cursor says errors and missing permissions blocked remote access on most affected machines, but malicious artifacts could remain. If you installed either version during that window, follow Cursor’s investigation and cleanup steps.
Claude’s watermark now reaches older models, but it can’t prove authorship
Anthropic expanded its invisible text watermark to Mythos 5, Opus 4.7, and Sonnet 4.6 on November 3. Detection remains a private preview, and a positive result can mean Claude merely edited the text. Sonnet’s mark can also become unreliable with changed sampling settings, so this isn’t a magic “you cheated” button.
ChatGPT can now read your iCloud Mail and Calendar
OpenAI added iCloud Mail and Calendar connections to ChatGPT, so it can find emails, check your schedule, and update calendar events after you connect them. They’re available on all ChatGPT plans on the web, desktop, and iOS. Your Apple inbox finally gets the same AI treatment as the rest of your schedule.
You can text your ChatGPT dot from iMessage or RCS
OpenAI is rolling out texting for ChatGPT dots, so you can message your always-on assistant through iMessage or RCS instead of opening ChatGPT. The beta is limited to some US Pro users, and the messages pass through a third-party provider, so they aren’t end-to-end encrypted between you and your dot.
NVIDIA’s Nemotron Transcribe lets you ask for the transcript you actually want
Nemotron 3.5 Transcribe can label speakers, translate speech, and follow instructions about whether to preserve every um or clean the transcript up. NVIDIA released downloadable weights supporting 40 languages and recordings up to an hour. It’s an offline model for recorded audio, so don’t mistake it for a ready-made live voice agent.
ChatGPT and Codex will share one usage budget in December
Starting in December, ChatGPT and Codex will draw from one token allowance on Plus, Pro, and Business. Demanding models and tasks eat more of it, so your coding agent and research session share the fridge. Instant chat stays outside the meter; OpenAI hasn’t published the allowance sizes in its announcement.
Apple opened Siri AI integrations to every developer
Apple’s Siri AI Developer Preview lets any developer expose app actions so people can book, order, or get things done through Siri. It’s available in the 27.3 developer betas; customer launches and App Store submissions come later. The useful part for builders is getting discovered through a request instead of another app icon.
Claude Message Threads stops agents uploading the same conversation over and over
Anthropic’s Message Threads beta lets agents send the conversation once, then only the new messages. Claude still reads the whole history and bills the same tokens; you save waiting time. Threads expire after five inactive minutes; Anthropic reports about 35% faster first responses on very long Claude Code requests.
Claude Code is pushing back on AI copies of other people’s games
Anthropic changed Claude Code’s Opus 5.5 instructions to steer it away from recreating someone else’s specific game. An Anthropic engineer says original games, genres, and projects built with your own designs and assets are still supposed to work. The change lands as game developers worry about AI-generated clones beating their own releases to market.