Gemini Nano Banana 2.1 Is GA: Half the Price, and a 23-Day Migration Window

The most consequential number in the October 6 release is not a benchmark. It is a date: October 29, 2026. That is when gemini-3.1-flash-image — the model everyone calls Nano Banana 2, and officially Gem​ini 3.1 Flash Image — stops being served. The replacement, Gem​ini Nano Banana 2.1, went generally available the same day the shutdown notice appeared, which leaves 23 days to change a model string in production. Pricing moved too, and it moved in the direction developers care about: Nano Banana 2.1 is billed at half of Nano Banana 2's output rate at every resolution, and its cheapest tier is cheaper than Nano Banana 2's cheapest tier even though 2.1 starts one resolution higher.

This is a GA release with a real migration deadline attached, not a point upgrade you can schedule at leisure. Everything below is what Google's own documentation and changelog say, separated from what only Google claims.

What actually shipped, in Google's own words

The Gemini API changelog entry for October 6, 2026 reads: "Gemini Nano Banana 2.1 generally available (GA): Released Gemini Nano Banana 2.1 (gemini-nano-banana-2.1), the latest high-efficiency image generation and conversational editing model." The same changelog, further down the same day, carries the second half of the story: "The gemini-3.1-flash-image model is deprecated and will be shut down on October 29, 2026. Migrate to gemini-nano-banana-2.1."

The documentation page for image generation positions it plainly: "An update to Nano Banana 2, serving as the primary high-efficiency workhorse model for image generation and conversational editing." Nano Banana 2's own row now reads "The previous-generation high-efficiency workhorse model," with the note that new projects should use 2.1. Google's model directory tags 2.1 with "New Stable" rather than preview.

• API identifier — gemini-nano-banana-2.1, stable; the outgoing one was gemini-3.1-flash-image.

• Output resolutions — 1K, 2K and 4K. The 512px tier that Nano Banana 2 offered is explicitly not supported on 2.1.

• Inputs — text, image, video or any combination; audio is not an input.

• Reference images — up to 14, specified as up to 10 high-fidelity object images and up to 4 images held for character consistency.

• Thinking mode — on by default and not switchable off, with minimal, medium and high levels; medium is the default.

• Grounding — Google Search grounding plus, new in this generation, Google Image Search grounding alongside web search.

• Watermarking — every generated image carries a SynthID watermark.

• Batch — available through the Batch API with up to a 24-hour turnaround, at half the interactive rate.

The same docs page also carries the list of things Google knows are limits: search grounding will not use real-world images of people found on the web, character resemblance caps at four characters and object fidelity at ten per workflow, and the model "won't always follow the exact number of image outputs that the user explicitly asks for." Those caveats are worth reading before you put 2.1 behind an automated pipeline that counts on a fixed number of returned images.

The price is the actual headline

Both models are token-billed with a per-image figure Google derives for you. Nano Banana 2.1's output is $30 per million image tokens, which Google's pricing page resolves to $0.0336 per 1K image, $0.0504 per 2K image and $0.0756 per 4K image at standard rates. Batch output is $15 per million tokens: $0.0168, $0.0252 and $0.0378 respectively. Input is $1.50 per million tokens whether it arrives as text, image or video, and text output is $7.50 per million tokens.

Nano Banana 2, by contrast, bills images at $60 per million tokens: $0.045 at 512px, $0.067 at 1K, $0.101 at 2K and $0.151 at 4K, with batch at $0.022/$0.034/$0.050/$0.076. Set the two side by side and the shape of the release becomes obvious.

• 1K image — Nano Banana 2.1 $0.0336; Nano Banana 2 $0.067. Half the price, same nominal resolution.

• 2K image — Nano Banana 2.1 $0.0504; Nano Banana 2 $0.101. Still half.

• 4K image — Nano Banana 2.1 $0.0756; Nano Banana 2 $0.151. Still half, and now a 4K render costs less than a 2K render did a generation ago.

• Cheapest possible render — Nano Banana 2.1's floor is $0.0336 at 1K, below Nano Banana 2's $0.045 floor at 512px, even though 2.1 has no 512px tier at all.

• Batch — $0.0168 per 1K image on 2.1, roughly a third of Nano Banana 2's standard $0.045 at its smallest size.

• Text and thinking output — 2.1 $7.50 per million tokens against Nano Banana 2's $3.00, which is the one line where the new model is the more expensive one; it only matters if you are extracting long text descriptions rather than images.

The one real capability regression is the dropped 512px tier. If a workflow deliberately traded resolution for the $0.045 floor, there is no drop-in cheaper tier here — 2.1's floor is 1K. In practice that is a price cut plus a resolution upgrade for the same or less money, but it is a change in the shape of the offer, not just the number.

The deadline nobody has time for

A shutdown date announced the same day as the replacement is the part of this release that will actually break things. Twenty-three days is enough to change an identifier and re-run an evaluation; it is not enough to re-tune prompts, re-shoot a reference library, or discover that your consistency rig depended on behaviour 2.1 does differently. The migration itself is genuinely small — the request shape, the 14-reference ceiling, the resolutions and the watermarking are all continuous with Nano Banana 2 — but the endpoint string is not, and a deprecated endpoint does not keep answering politely after the date.

Two practical notes. First, the deprecation applies to the API model. Anything you built directly against gemini-3.1-flash-image keeps working until the cutoff and then stops; the successor is a different identifier, not an alias. Second, a 23-day window is exactly the case for not having a single hard-coded model string in the critical path. If your image generation goes through a routing layer rather than a vendor SDK, the identifier is configuration, and the failure mode of a missed migration is a config change rather than a deploy.

That is the argument for a gateway on its own merits, independent of price. OrcaRouter fronts 200+ models behind one OpenAI-compatible endpoint, passes provider list price through with 0% markup, and moves a request to another provider automatically when one starts failing. Image models in that catalogue include google/gemini-3.1-flash-image-preview and openai/gpt-image-2, so the sensible comparison while you evaluate 2.1 is a live one rather than a paper one. We do not route gemini-nano-banana-2.1, and we will not claim to — the model is available from the vendor's own API and from several third-party platforms.

What is vendor-reported, and what nobody has measured yet

Google's changelog claims "significant improvements" in visual quality, prompt adherence, multi-turn character consistency and text rendering, plus wide and panoramic aspect-ratio generation. Every one of those is a vendor claim about its own model, unreproduced as far as the public record shows. The docs page adds that Gemini Nano Banana 2.1 and Gemini 3.1 Flash Image add Google Image Search grounding alongside web search, which is a feature statement rather than a quality statement and can be checked directly by anyone with an API key.

What we did not find is more interesting than what we did. There is no independent arena placement for Gemini Nano Banana 2.1 — no Elo, no vote count, no third-party render comparison — in the days after release. For a model in a family that has been benchmarked heavily, and where the sibling entries have thousands of human votes behind them, the absence is the honest state of the evidence on October 7. The improvement claims are Google's; the price and the deprecation date are facts you can verify in the changelog.

Who should move now

If you are already calling gemini-3.1-flash-image in production, there is no decision to make: migrate this month, before the deadline makes it for you. The change is a model identifier, the price is roughly halved, and the one regression — no 512px tier — affects only workflows that deliberately shipped sub-1K images.

If you are choosing an image model from scratch, the interesting question is no longer "is 2.1 better than 2." It is whether a 1K render at $0.0336 makes Nano Banana 2.1 competitive against the cheaper half of the market rather than the premium half it used to sit in. That is a comparison against Qwen Image 3.0 and the open-weight tier, not against Nano Banana Pro, and it is worth running on your own prompts rather than on either vendor's slides.

If you are on Nano Banana 2 Lite, nothing here forces your hand; Lite remains the cheapest and fastest tier, and 2.1 is priced and positioned well above it.

What to watch next

Two things would change this picture. The first is an independent placement: the moment a board with human votes carries Gemini Nano Banana 2.1, the vendor's improvement claim either survives contact or does not. The second is whether the October 29 shutdown slips — Google has moved such dates before, and a delay would tell you how much of the ecosystem actually migrated in three weeks.

Until one of those lands, treat this as what it is: a half-price generation with a hard deadline, a real but unmeasured quality claim, and a migration small enough that there is no good reason to be running gemini-3.1-flash-image on October 30.