Google released Nano Banana 2.1 on October 6 as a generally available image model. Developers now have a stable model ID, output controls and published API prices. Google's quality comparisons remain its own tests.

Google's October 6 Gemini API release notes mark Nano Banana 2.1 as generally available and name gemini-nano-banana-2.1 as its stable API ID. The company recommends it for new work in place of Nano Banana 2, gemini-3.1-flash-image, which it has deprecated without announcing a shutdown date. For someone deciding whether to use the new model, the practical change is a documented route from a text or reference-image input to a 1K, 2K or 4K image, with a published price and explicit limits.

The big change

What changed: Google released Nano Banana 2.1 as a generally available image generation and conversational editing model on October 6, 2026. Its model card says it is based on Gemini 3.6 Flash and lists the Gemini app, AI Studio, Gemini API, Search AI Mode, Ads, Flow and Stitch as distribution channels.

How it works for a reader: In the API, a developer selects gemini-nano-banana-2.1, supplies text and optionally an image, then reads image and text output. The image generation guide documents up to 14 reference images, output size and aspect-ratio controls, optional search grounding, and three thinking levels. A creator can instead check the model in Google AI Studio or another listed Google product; the card's channel list does not establish that every account has identical access or controls.

Why it matters: The paid API price for a standard 1K output image is $0.0336, compared with $0.067 for Nano Banana 2 on Google's price page. That is a price comparison for the specified API output, not a measured comparison of image quality or latency. BIG CHANGE reviewed documentation and did not run the model.

Choose an interface, then check the output

For a creator who wants to make or revise an image without building an integration, Google's model card names the Gemini app and Flow among its channels, while the developer model page links directly to Google AI Studio. Availability and product controls should be checked in the intended account before committing to a workflow.

For an application, the API guide's current examples use the Interactions API with the stable model ID. A text prompt can return image data; an editing request adds image data and an instruction. The model page lists text, images, video and PDF as supported inputs, and image and text as outputs. Audio input is unsupported. The guide says the default response includes both image and text, while response_format can request images alone or specify an aspect ratio and image size. The default image size is 1K; 2K and 4K are available. The developer model page lists a 131,072-token input limit and a 32,768-token output limit. Its table does not support caching, function calling, structured outputs or the Live API.

Reference images can help with editing and compositing. Google says the model accepts up to 14 references in a workflow, with character resemblance for up to four characters and fidelity for up to ten objects. Those numbers describe supported inputs and design aims, not a guarantee that every subject will remain unchanged. Google also offers minimal, medium and high thinking settings, with medium as the default. The guide shows Google Web Search grounding and an optional Image Search type; an application must configure the search tool to request it. Google says web image search does not support using real-world images of people in this image-generation path.

Before adopting it, check that the intended account exposes the interface, use the documented model ID and input type, and inspect a returned image at the size it will be used. BIG CHANGE has not completed that generation; this article maps Google's documentation.

Price the actual request

Google's pricing table has no free API tier for this model. Standard paid rates are $1.50 per million input tokens for text, image or video; $7.50 per million output text and thinking tokens; and $30 per million output image tokens. Google equates that image rate to $0.0336 at 1K, $0.0504 at 2K and $0.113 at 4K. Its Batch rates are half the corresponding token rates, with listed image equivalents of $0.0168, $0.0252 and $0.0567. Search grounding has a shared allowance of 5,000 requests per month across Gemini 3.x models, then $14 per 1,000 requests. Google warns that one user request can trigger multiple billed search queries. Check the live pricing page and account billing setup before deploying a high-volume workflow.

Google's results and the remaining gaps

Google reports better visual design, editing, text rendering and subject consistency than earlier models. In its October 2026 model card, the company describes human side-by-side preference evaluations and a single-sided automated factuality rating, combining public benchmarks with internal sets. Its infographic factuality score for Nano Banana 2.1 with thinking is 0.521, versus 0.179 for Nano Banana 2 with thinking in that table. Those are Google's evaluation results under its selected conditions. Decrypt's launch coverage also identified the absence of independent testing at launch; it does not verify the scores.

The model card itself lists blurry small text, imperfect character consistency, partial instruction following in masked edits, left/right confusion and remaining limits in factuality and spatial reasoning. It also warns of hallucinations and occasional slow or timed-out requests. A search-grounded image therefore still needs fact checking, and a production image still needs inspection at its intended display size. All generated images include Google's SynthID watermark, according to the API guide.

One documentation conflict deserves attention during integration: the model card describes a context window of up to one million tokens and larger text output, while the API model page lists 131,072 input and 32,768 output tokens. The model page is the explicit API specification used above. Google should reconcile the two figures; developers should use the live endpoint documentation and their account limits when sizing requests.

Sources & further reading

  • Google DeepMind: Nano Banana 2.1 model card (published October 2026; model-card index updated October 6). Describes the model basis, distribution, Google's evaluation approach and known limitations. Its token figures differ from the developer model page.
  • Google AI for Developers: Nano Banana 2.1 model page (updated October 6, 2026). Gives the stable API ID, supported modalities, API token limits, defaults and unsupported capabilities.
  • Google AI for Developers: image generation guide (consulted October 7, 2026). Documents current Interactions API examples, reference images, grounding, output controls and limitations. Examples describe an interface; BIG CHANGE did not execute them.
  • Google AI for Developers: release notes (October 6, 2026 entry). Establishes the general-availability date and Nano Banana 2 deprecation without a shutdown date.
  • Google AI for Developers: pricing (consulted October 7, 2026). Lists paid standard and Batch token rates, image equivalents and search-grounding charges. Prices can change.
  • Decrypt: launch coverage (October 6, 2026). Independent report of launch context that notes the lack of independent testing at launch; its performance discussion relies on Google's results.