Gemini Nano Banana 2.1 is Google's new generally available image generation and editing model in the Gemini API. It replaces Nano Banana 2, which shuts down on 29 October 2026. Teams with an image feature built on the older model have a short, fixed window to test the new one, update their prompts and switch.
What was released
Gemini Nano Banana 2.1 was announced by Google on 6 October 2026 in the Gemini API release notes, as a generally available model with the ID gemini-nano-banana-2.1. It is an update to Nano Banana 2 (gemini-3.1-flash-image) and sits alongside Nano Banana Pro, which Google keeps as the higher-end option.
The same note deprecates the model it replaces. According to the Gemini deprecations page, gemini-3.1-flash-image shuts down on 29 October 2026, with gemini-nano-banana-2.1 as the recommended replacement.
What actually changed
The Nano Banana 2.1 model page and the image generation guide list these differences from Nano Banana 2:
- Better output where it usually breaks. Google reports improved visual quality, prompt adherence, text rendering and infographic layout, and more consistent characters across a multi-turn editing session.
- Wide and panoramic images that hold together. Ratios of 1:4, 4:1, 1:8 and 8:1 are supported, and the model page says tiling artefacts on these ratios at 2K and 4K are fixed.
- More reference images. One request can combine up to 14 reference images: up to 10 objects and up to 4 characters to keep consistent.
- Search grounding and thinking. The model can ground images with Google Web and Image Search, and has configurable thinking levels (minimal, medium by default, and high).
- No smaller size. Output is 1K by default, with 2K and 4K available. The 512px option of Nano Banana 2 is not supported on 2.1.
- A different price shape. Image output is cheaper: on the Gemini API pricing page, Nano Banana 2.1 lists about half the Nano Banana 2 price per image at 1K, 2K and 4K. Input tokens and text and thinking output cost more per token than before, the Batch API halves the rates, and there is no free tier.
Inputs can be text, images, video or PDF, and the model returns images and text. The model page also lists what it does not support: function calling, structured outputs, context caching, code execution and URL context.
What it means for teams building AI products and agents
In CodeDTX's view, the deadline matters more than the quality gains. Here is how we would handle it.
- Find every call to the old model ID now. Search code, configuration, prompt stores and any agent tool definitions for
gemini-3.1-flash-image. Calls that are still pointing at it after the shutdown date will fail. Our guide to what to do when a model is deprecated covers the inventory and the fallback plan. - Re-run your image prompts, not just a smoke test. Better prompt adherence also means different output. Brand templates, text-heavy layouts and character sheets built on the old model should be checked side by side, because a prompt tuned for the old model can now produce a layout your team did not expect.
- Recalculate cost per image. If most of your spend is generated images, the new per-image rates should lower it. If you send long prompts, many reference images or video and use high thinking, the higher input and thinking prices can cancel that out. Teams relying on 512px output will pay for 1K images instead. Measure cost per finished asset, as in controlling what an AI agent costs to run.
- Keep image generation behind a tool your agent calls. Nano Banana 2.1 does not do function calling or structured outputs, so it should not be the agent's main model. Let a text model plan the work and call image generation as a tool with a fixed input and output contract; see how to design tools an AI agent can use well. Swapping model IDs then touches one place.
- Review grounded images before they ship. Search grounding pulls in web and image results, which raises questions about what reached the final image. For anything public, keep a person approving the output and log the prompt and sources, as in how to audit what an AI agent did. Google says every generated image carries a SynthID watermark.
When not to switch yet
There is no option to stay on Nano Banana 2 past the shutdown date, so the real choice is between Nano Banana 2.1 and another model. If your workload needs studio-grade layouts that 2.1 still cannot match in your tests, compare it with Nano Banana Pro rather than forcing the switch. If you only need 1K images and send long prompts, look at Nano Banana 2 Lite (gemini-3.1-flash-lite-image): its pricing matches 2.1 per 1K image with lower input and thinking rates, and it has no shutdown date announced. And do not turn on search grounding or high thinking by default just because they are available: add them only where your evaluation shows a gain worth the extra cost and latency.
Frequently asked questions
What is Gemini Nano Banana 2.1?
Gemini Nano Banana 2.1 is Google's image generation and conversational editing model, released as generally available in the Gemini API on 6 October 2026 with the model ID gemini-nano-banana-2.1. It updates Nano Banana 2 with better image quality, prompt adherence, text rendering and character consistency, plus wide and panoramic aspect ratios, while keeping the speed and cost profile of the Flash models.
When does Nano Banana 2 shut down?
Google has deprecated gemini-3.1-flash-image, known as Nano Banana 2, and says it will shut down on 29 October 2026. The recommended replacement is gemini-nano-banana-2.1. Any application or agent still calling the old model ID after that date will start returning errors, so teams should find those calls now, test the new model on real prompts and update the configuration before the deadline.
Is Nano Banana 2.1 cheaper than Nano Banana 2?
For generated images, yes: Google's pricing page lists lower per-image prices at 1K, 2K and 4K than for Nano Banana 2. Input tokens and text and thinking output cost more per token, though, and the 512px size is not available. Whether your total bill falls depends on how long your prompts are, how many reference images you send and which thinking level you use.
Can an AI agent use Nano Banana 2.1 directly?
Not as its main model. The model page lists function calling and structured outputs as unsupported, so it cannot plan steps or call your tools. The usual pattern is a text model that plans the task and calls image generation as one of its tools, with clear inputs such as the prompt, reference images, aspect ratio and size, and a person reviewing images before they are published.
Do Nano Banana 2.1 images carry a watermark?
Yes. Google's image generation guide says all images generated by its Gemini image models include a SynthID watermark, which is an invisible marker that identifies the image as AI generated. Teams publishing these images should still follow their own disclosure rules and keep a record of the prompt and any search sources used, so they can answer questions about where an image came from.
