Last updated: October 3, 2026
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.