On October 6, Google released Nano Banana 2.1, the model ID gemini-nano-banana-2.1, as a generally available image generation and editing model. The headline number spread fast: a generated image through the Gemini API now costs about half what Nano Banana 2 charged. A 1K image is $0.0336 instead of $0.067. A 2K image is $0.0504 instead of $0.101.
That is a real price cut, and for most creative teams it is not the number that matters. At these prices the image itself is close to free. What costs money is the designer who has to fix a blurry price, a drifting product label, or a headline the model wrote in a language nobody on the team reads. Google’s own documentation tells you where those fixes will come from, if you read past the pricing table.
A note on evidence. This is a primary-source review. I have not run Nano Banana 2.1 on a paid Gemini API project or inside a Google Ads account for this piece. Everything below comes from Google’s Gemini API release notes, model page, image generation guide, pricing page, and deprecations page, the Google DeepMind model card, and the Google Ads Help page for Asset studio, all checked on October 7 and 8. I cross-checked against Android Authority, Vercel’s changelog, Apidog’s API guide and model comparison, TechCity Authority, and Mohammad Nasravi’s AI Daily. Where Google’s own pages disagree, I say so.
What it is
Nano Banana is Google’s name for Gemini’s native image generation. Version 2.1 is an update to Nano Banana 2 (Gemini 3.1 Flash Image), which Google now calls “the previous-generation high-efficiency workhorse.” The model card says 2.1 is built on Gemini 3.6 Flash. Google positions it as the efficient counterpart to Nano Banana Pro, which stays the premium tier.
It takes text, images, video, and PDFs as input and returns images and text. It can combine up to 14 reference images, with character consistency for up to 4 characters and object fidelity for up to 10 objects. It can ground an image in Google Web Search and Google Image Search. It outputs 1K, 2K, or 4K, and every image carries a SynthID watermark.
The distribution list is what makes this a marketing story rather than a developer one. The model card lists the Gemini app, Google AI Studio, the Gemini API, AI Mode in Search, Google Ads, Google Flow, and Google Stitch. Android Authority’s launch report adds Gemini Enterprise. The same model your agency calls through the API is also, per Google, landing in the tool your media buyer generates assets in.
What changed on October 6
The price per image halved at 1K and 2K. Google’s pricing page lists image output at $30 per million tokens, down from $60 on Nano Banana 2. A 1K image uses 1,120 tokens and a 2K image 1,680, so the per-image prices fall by half. The Batch API halves them again: $0.0168 at 1K and $0.0252 at 2K, for jobs that can wait up to 24 hours.
Thinking got heavier by default. Gemini 3 image models reason before they draw, and the image guide says this “cannot be disabled in the API.” Nano Banana 2.1 adds a third thinking level and defaults to it: minimal, medium (default), and high. Nano Banana 2 defaulted to minimal. If your pipeline swaps the model ID and changes nothing else, it now thinks more on every request, and the guide notes that “thinking tokens are billed by default.”
Wide banners stopped tiling. The model page says 2.1 “fixed tiling artifacts” on the 1:4, 4:1, 1:8, and 8:1 aspect ratios. Those shapes map to skyscrapers, leaderboards, and site header strips. The fix is listed at 2K and 4K only.
The old workhorse was deprecated. The release notes mark gemini-3.1-flash-image as deprecated and tell developers to migrate. The deprecations page lists no shutdown date yet, so this is a planning signal, not a deadline.
What works
Google published numbers against its own older models. The DeepMind model card reports side-by-side human preference scores. On overall text-to-image preference, 2.1 with thinking scores 1050, against 990 for Nano Banana 2 and 935 for Nano Banana Pro. Multi-character consistency is 1106 against 978 for Nano Banana 2. Infographic factuality, an auto-rated score, is 0.521 against 0.179. These are Google’s evaluations of Google’s models, so treat them as a reason to test, not a result. Apidog’s launch-week comparison said Google had published no benchmarks at all, so make sure whoever briefs your team has read the current card.
Consistency is aimed at the jobs marketers actually run. Four characters and ten objects held across one job covers a campaign mascot, a spokesperson, and a product line in a single scene. The model card’s editing scores include product consistency and multi-reference editing, and Vercel’s launch note leads with “product recontextualization,” meaning placing an existing product in a new scene. That is the catalog and seasonal refresh work most teams outsource or skip.
Localization is a stated use. The model card lists “localized text rendering across several languages” as an intended use, and Google’s own guide demonstrates translating an infographic into Spanish in a second turn with the instruction to change nothing else. For a team running one campaign in six markets, that is the expensive part of the job.
At volume, the bill is small and predictable. Apidog’s worked example puts 1,000 product images at 2K from short prompts at about $50.55 on 2.1, against about $101 on Nano Banana 2. Through Batch, the output side drops to $25.20.
Where it breaks or is fenced
The cheap default is where the small print blurs. The model card’s known limitations open with “poor text rendering in small text (often blurry in 1k model), long paragraphs, page length.” The model page says the default output resolution is 1K. Put those together and the size that headlines the price cut is the one Google warns against for prices, disclaimers, ingredient lists, and pack copy. Google’s own prompting guide still says to use Nano Banana Pro “for professional asset production” when the job is text. Mohammad Nasravi, an in-house AI creative director, put the practical version in his AI Daily brief: “Do not trust the small print on a pack: composite the real label in post.”
Input got three times more expensive. Input tokens cost $1.50 per million on 2.1 against $0.50 on Nano Banana 2. Text and thinking output cost $7.50 per million against $3. Apidog ran the math on a request with 14 reference images and a 1K output: 2.1 still wins, but the saving shrinks from 50 percent to roughly 23 percent. Long multi-turn edit sessions, where earlier turns ride along as context, narrow it further.
4K got a smaller cut than the coverage says. Launch-day guides from Apidog and TechCity Authority quoted a 4K image at about $0.076, based on 2,520 tokens. Google’s pricing page, last updated October 7, now says a 4K image on 2.1 consumes 3,780 tokens and costs $0.113. That is a 25 percent cut from Nano Banana 2’s $0.151, not 50. If someone on your team built a budget from a launch-day blog post, rebuild it from the live page.
You lose the draft tier and some knobs. Nano Banana 2.1 does not support the 512px size Nano Banana 2 offered. The model page lists no context caching, no function calling, no structured outputs, and no Flex or Priority inference. There is no free API tier. Batch is supported.
Inside Google Ads, you cannot see the model. The model card says 2.1 is distributed through Google Ads. The Google Ads Help page for Asset studio, as of October 8, still says Asset studio was upgraded with “the Nano Banana Pro image generation model.” TechCity Authority, reporting on Flow, noted that a model menu “is just a label” and that early testers could not tell whether requests were really routed to the new model. If you generate ad images in Google’s tools, the model under you can change without a line in your change log.
Google’s pages disagree on dates. The pricing page says the original Nano Banana (gemini-2.5-flash-image) “will be shut down on October 2, 2026.” The deprecations page lists its shutdown as March 15, 2027. If anything in your stack still calls that model, check which one is true for your project before you plan around either.
The rest of the limits list matters too. The card names imperfect character consistency between input and output, partial instruction following in masked edits, occasional left and right confusion, and limited 3D reasoning. Left and right confusion matters if your product has a front.
Who it is for right now
It fits best for in-house creative ops and agency production teams who already run image generation through the API or a gateway, produce at volume, and have someone who can change a pipeline setting. The jobs that pay off are product recontextualization across a catalog, seasonal refreshes, mascot or spokesperson scenes that have to stay on model, wide banner sets at 2K, and localization where a designer still checks the final file.
It fits less well for anyone who wants finished, text-heavy ads straight out of the model, for teams whose only access is through Google Ads or Flow and who cannot pin or verify the model version, and for pipelines that depend on 512px drafts.
What it costs
Through the Gemini API on the paid tier, Google’s pricing page lists Nano Banana 2.1 at $1.50 per million input tokens, $7.50 per million text and thinking output tokens, and $30 per million image output tokens. Per image, that is $0.0336 at 1K, $0.0504 at 2K, and $0.113 at 4K. Batch halves each: $0.0168, $0.0252, and $0.0567. Grounding with Google Search includes 5,000 free requests a month shared across Gemini 3.x models, then $14 per 1,000.
For comparison, Nano Banana 2 is $0.067, $0.101, and $0.151, and Nano Banana Pro is $0.134 at 1K or 2K and $0.24 at 4K.
Here is the arithmetic that should change how you think about it. A campaign set of 40 assets, localized into six languages at 2K, is 240 images. On 2.1 that is about $12 in image output. Even if you generate five candidates for every final asset, you are near $60. The model is not the budget line. The review is. Google does not publish a per-image price for 2.1 inside Google Ads, Flow, or the Gemini app.
What to do this week
Treat October 6 as a cheaper model with a different default, not a free upgrade.
- Set 2K as the floor for anything with type on it. Leave 1K for mood boards, concept rounds, and backgrounds where no one will read anything.
- Keep live copy out of the generation. Prices, offers, legal lines, and pack text go on as an overlay in your design tool, so a model swap can never change what your ad says.
- Pin the thinking level on purpose. Run your standard prompt set at minimal and at the new medium default, and compare cost and fix time before you accept the default.
- Run one side-by-side before you switch. Take one product with a hard pack, small print, and a person in frame. Run the same prompts through Nano Banana 2 and 2.1 at 2K, and log where the label breaks and how long each fix takes.
- Rebuild your budget from the live pricing page, especially at 4K, and include input and thinking tokens for reference-heavy jobs.
- Ask your Google rep which model Asset studio uses in your account today, and add a quick visual check on generated assets whenever Google announces a model change.
- Find any calls to the original Nano Banana model in your stack and move them, whichever shutdown date turns out to be right.
Nano Banana 2.1 makes generated images cheap enough that nobody should argue about the per-image price again. That shifts the work to the parts Google flags in its own fine print: small text, consistency at the edges, and a model that can change under your ad tools without notice. Spend this week’s time there, not on the price table.
