Nano Banana 2.1 Is Out: Google Retired the Model Above It and Halved the Price

Two sentences in the October 6, 2026 release note are hard to hold together. The first puts Nano Banana 2.1 into general availability under the model id gemini-nano-banana-2.1. The second announces that Nano Banana 2 — the model it supersedes, shipped as Gemini 3.1 Flash Image — will be shut down on October 29, 2026. Then the rate card does something stranger: Nano Banana 2.1 is listed at $0.0336 per 1K image, which is half of what Nano Banana 2 charges for the same output and exactly what the budget tier, Nano Banana 2 Lite, charges. The vendor's high-efficiency workhorse got better, cheaper and, in 22 days, the only one of the two that still answers.

So this is a launch that actually happened, not a leak: there is a published model card on ai.google.dev, a changelog entry, and a post from Google's own account describing the same model in the same words. What is missing is the one thing this blog would normally open with — a number. There is no entry for Nano Banana 2.1 on Artificial Analysis's image arena, no third-party evaluation of any kind, and therefore no head-to-head against GPT-Image-2.5, which is the comparison the first wave of reactions asked for before anything else. Every quality claim below is Google's own and is labelled as such. That gap is not a research failure on our side; it is the state of the record on day one.

What Google says it shipped

Nano Banana 2.1 is framed as an update to Nano Banana 2 rather than a new tier, and the documented capability list reads like a set of repairs to that model's weakest points:

• Resolutions — 1K, 2K and 4K, defaulting to 1K. It does not offer the 512px (0.5K) size that Nano Banana 2 introduced, so the cheapest thumbnail tier disappears with the upgrade.

• Wide and panoramic frames — the model fixes tiling artifacts on the 1:4, 4:1, 1:8 and 8:1 aspect ratios at 2K and 4K. Those extreme shapes are exactly where a seam shows up, and it is the most concrete claim on the card. Fourteen ratios are supported in total.

• Multi-image fusion — up to 14 reference images in a single workflow, with character consistency for up to four characters and object fidelity for up to ten. Nano Banana 2 Lite supports the 14 object references but no character-consistency references at all.

• Grounding — Google Web Search plus, new for this generation, Google Image Search grounding. Google's documentation adds a caveat that matters for anyone building around people: search grounding cannot use real-world photographs of people pulled from the web.

• Thinking levels — minimal, medium and high, with medium as the default. Nano Banana 2 and Nano Banana 2 Lite expose only minimal and high, so medium is genuinely new control surface, not a relabel.

• Input and limits — text, image, video and PDF in; a 131,072-token input limit and a 32,768-token output limit. Video-to-image generation is supported.

• Provenance — every generated image carries a SynthID watermark.

The model card is equally clear about what Nano Banana 2.1 does not do: no function calling, no context caching, no structured outputs, no URL context, no Live API. It does support the Batch API, which becomes important in the next section.

The rate card is the actual story

Read the three tiers side by side and the release stops looking routine. Google bills image output per million tokens and the token cost of an image scales with its resolution, so the same model quoted as "$30 per million image tokens" means three different per-image rates at three sizes.

• Nano Banana 2.1 — $1.50 per million input tokens and $30.00 per million image-output tokens, which Google resolves to $0.0336 per 1K image, $0.0504 at 2K and $0.0756 at 4K. Batch halves all of it, to $0.0168, $0.0252 and $0.0378.

• Nano Banana 2 (Gemini 3.1 Flash Image) — $0.50 input and $60.00 per million image tokens, or $0.067 at 1K, $0.101 at 2K and $0.151 at 4K. Batch halves that to $0.034, $0.050 and $0.076.

• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) — $0.25 input and $30.00 per million image tokens, or $0.0336 for a 1K image, $0.0168 in batch. It produces 1K only; 2K and 4K are not supported.

Put those next to each other and the pricing logic inverts. Nano Banana 2.1's standard 1K rate of $0.0336 undercuts Nano Banana 2's batch rate of $0.034 — the new model at full price is cheaper than the old model at half price, for an image Google describes as an improvement. It is also the identical rate to Nano Banana 2 Lite, which means the two cheapest ways to buy a Google image are now the same price and no longer the same product. The differentiator has moved off the cheque and onto the ceiling: at the same money, one model stops at 1K and the other reaches 4K, takes video as context, grounds on search results and holds more references.

The premium tier is untouched. Nano Banana Pro (Gemini 3 Pro Image) remains the top of the family for complex assets, and Google's release note positions Nano Banana 2.1 explicitly as "the more efficient counterpart" to it rather than a replacement.

October 29 is the date on the calendar

The part of this release with real operational teeth is the deprecation, and Google's own deprecation table is unambiguous. gemini-3.1-flash-image — Nano Banana 2 — becomes unavailable on October 29, 2026, and the documented migration target is gemini-nano-banana-2.1. The same table shows how thoroughly Google is consolidating onto the new id: gemini-2.5-flash-image was already shut down on October 2, 2026, and the entire Imagen 4 line — the three imagen-4.0-* models — went dark on August 17, 2026. Every one of those rows now points at Nano Banana 2.1.

Twenty-two days is not a lot of runway for a production pipeline, and the shape of the migration is unusually favourable: a 1:1 identifier swap that, on the pricing above, reduces the bill. The awkward cases are the ones that depend on behaviour nobody has measured. If your prompts were tuned against Nano Banana 2's specific failure modes, you are moving to a model with no third-party profile, on a deadline, with no option to wait for one.

Why there is no benchmark, and why that is the real caveat

The temptation with a 0.1 release is to assume it is a checkpoint swap. Google's language argues otherwise — "significant improvements in visual quality, prompt adherence, multi-turn character consistency, text rendering" is a longer list of claims than a maintenance bump usually carries — but that sentence is the vendor describing its own model, and it has not been reproduced by anyone outside Google. There is no vendor eval chart in the model card, no vendor preference study, and no independent board entry.

For scale, the model it replaces was measured, modestly and without controversy. In our own last reading of Artificial Analysis's image arena on October 6, 2026, Nano Banana 2 sat in the top handful of the text-to-image board at roughly Elo 1,125, with Nano Banana Pro a little behind it and Nano Banana 2 Lite lower still. Those are votes from people choosing between two anonymous images, and they are the kind of evidence a 2.1 has to move. Nothing on the board has moved yet for Nano Banana 2.1 because the board has not seen it. That is the honest ceiling on everything this article can tell you about quality: Google says the wide-format tiling is fixed and the text rendering improved, and nobody outside the company has checked.

There is a specific risk worth naming. The predecessor Nano Banana 2 Lite is cheap, fast and capped at 1K, and Google has just published a model that matches it on price while claiming a much higher ceiling. If the claims hold, the Lite tier exists only for people who need sub-two-second latency, and Google's product line has quietly collapsed from four image models to about two and a half. If the claims do not hold — if 2.1's outputs are indistinguishable from its predecessor's, as the earliest eyeball reports about the interim Flow label suggested before launch — then the price is the whole release. Only blind votes will separate those two worlds, and there are none.

What is actually actionable this week

Two things, and neither involves waiting for a benchmark. The first is the migration itself: if anything in your stack calls gemini-3.1-flash-image, it has an expiry date of October 29 and a documented replacement that costs less. Do the swap early, while you can still run both, rather than on the week the old endpoint disappears.

The second is about how the swap is wired. The reason a Google price list matters to a routing layer is the same reason a deadline does: on OrcaRouter, each provider's list price is passed through with 0% markup, so a change on Google's rate card — the halving in this release included — is live on our side the same day rather than at the next renewal. Our catalogue carries the Google image models that exist as callable routes today, including google/gemini-3.1-flash-image-preview (the preview snapshot of Nano Banana 2), google/gemini-3-pro-image-preview for Nano Banana Pro, and google/gemini-2.5-flash-image. To be exact about the state of things rather than convenient: Nano Banana 2.1 itself is not on OrcaRouter yet, because it reached general availability on October 6 and stable-API routes are added as they settle, and neither is Nano Banana 2 Lite. What one API across more than 200 models buys you in the meantime is that the migration is a model slug in a request, and the day the new route lands it is the same kind of change — no second contract, no second billing relationship, no second SDK to learn while a shutdown clock runs.

If you would rather not commit a production path to an unmeasured model at all, the failover behaviour is the honest answer to that too: routing the same job through more than one provider means an underperforming model degrades a request instead of breaking it, which is a reasonable posture for the first few weeks of a model nobody has scored.

The question this release leaves open

Google shipped a cheaper, wider, better-documented image model and retired the one above it, and the version of this story that gets written in a month depends entirely on a leaderboard that does not exist yet. Three things would settle it, in descending order of usefulness. An Artificial Analysis entry, because that turns "significant improvements" into a number and tells you whether the Lite tier still has a reason to exist. A published model card from Google DeepMind with evaluation detail, because it is the difference between a vendor claim and a vendor measurement. And a first independent head-to-head against GPT-Image-2.5, because that is the comparison everyone actually has in mind and the one no one can currently make.

Until then the factual summary is short. Nano Banana 2.1 is generally available as of October 6, 2026, it is the cheapest full-resolution image model Google has ever sold, it is the mandatory destination for three deprecated model lines, and it has not been scored by anyone with no stake in the outcome. Act on the deprecation. Wait on the quality claim.