Nano Banana 2.1 Is Out: Older Image Model Ends October 29

Google released Nano Banana 2.1, its new model for generating and editing images, on October 6, 2026. On the API, each image costs half what it did with the previous model, and that previous model shuts down on October 29. It matters to anyone generating images with Gemini from a website, a plugin or an automation.

What we know

What it is and when it shipped: the Gemini API changelog has an entry dated October 6, 2026 announcing general availability of Gemini Nano Banana 2.1, with the model ID "gemini-nano-banana-2.1".
What improves, according to Google: visual quality, prompt adherence, character consistency across turns, text inside the image, and wide and panoramic formats (1:4, 4:1, 1:8 and 8:1), at 1K, 2K and 4K.
What shuts down: the same entry deprecates "gemini-3.1-flash-image" (Nano Banana 2) and sets its shutdown for October 29, 2026. Google's deprecations page repeats the date and names the replacement: "gemini-nano-banana-2.1".
What an image costs: the API pricing page lists $0.0336 per image at 1K, $0.0504 at 2K and $0.0756 at 4K. With the previous model it was $0.067, $0.101 and $0.151. Neither model has a free tier on the API.
What goes up: input rises from $0.50 to $1.50 per million tokens, and text and thinking output from $3 to $7.50, according to the same page.
Where it is: the Google DeepMind model card, published October 6, lists the Gemini app, Google AI Studio, the Gemini API, AI Mode in Search, Google Ads, Google Flow and Google Stitch. It gives no countries or plans.
In Search: Search Engine Roundtable reported on October 6 that Google and Robby Stein, a product lead for Search, said on X that the model is rolling out in AI Mode.
Watermark: every generated image includes a SynthID watermark, according to Google's image generation guide.

What changes and what doesn't

The reference model changes for API users. Google's guide recommends Nano Banana 2.1 for all new projects, and the previous model ID shuts down on October 29. That is 23 days after release.

The math changes. A thousand 1K images go from about $67 to $33.60: half. That is this article's estimate, and it counts image output only. Input costs three times as much. And the model's thinking, which the guide says is on by default and cannot be disabled in the API, is billed as text output. Anyone sending many reference images or long prompts will save less.

The minimum size changes. The size table for Nano Banana 2.1 starts at 1K. The previous model's table also had 512 pixels.

The rest stays put. Nano Banana 2 Lite and Nano Banana Pro have no announced shutdown date. The original Nano Banana shuts down on March 15, 2027.

The quality gains are Google's own measurements. Its model card gives Nano Banana 2.1 an overall preference score of 1050, against 990 for the previous model, in its own side-by-side human evaluations. No independent tests were reviewed for this story.

The card also admits limits: blurry small text at 1K, left-right confusion and factual errors. On the infographic factuality test the model scores 0.521 with thinking; the previous one, 0.179. The card does not explain the scale.

How to tell if it affects you

It affects you if your website, your store or an automation generates images through the Gemini API. If you only use the Gemini app, the shutdown date requires nothing from you. Seven steps:

Search for the old model ID. Check your code, plugin settings and automations for "gemini-3.1-flash-image". Careful: "gemini-3.1-flash-lite-image" is a different model and is not shutting down. Search for the exact name.
If someone else built it, ask. Your developer or the plugin vendor: "Which Gemini image model does my site use, and will you switch it before October 29?"
Switch and test before the date. Replace the ID with "gemini-nano-banana-2.1" and generate your usual images. It is a different model: results may vary.
If you requested 512-pixel images, the minimum is now 1K. Resize the image on your server before publishing it.
Redo the math with your real usage. Look at your billing for how many input tokens you spend per image: that part triples.
Before publishing an AI image on a business site, check what the card itself flags as weak: small text, the figures and facts in an infographic, and whether the product looks like the real one. Then compress it and add alt text.
Do not present it as a real photo. It carries SynthID, the watermark Google adds to images its models generate.

Prompts do not have to be in English: the guide's list of best-performing languages includes Spanish, German, French, Portuguese, Japanese and others.

Sources

Updates: this note will be extended if Google changes the shutdown date or publishes its own post about the model.