Nano Banana 2.1: half the price, and the part Pro still wins

What Nano Banana 2.1 is

Nano Banana 2.1 is Google's Flash-tier image generation and editing model, released on 6 October 2026 under the API identifier gemini-nano-banana-2.1. It is built on Gemini 3.6 Flash and replaces Nano Banana 2 — the model sold to developers as gemini-3.1-flash-image — which had launched in February 2026.

The naming is the first thing to get straight, because the family has grown into five models that share two naming systems:

Google's own description is careful and worth repeating exactly: it calls 2.1 the more efficient counterpart to Pro. That is not a synonym for better. It means this release is a cost and speed play with a genuine capability upgrade attached, and the upgrade is not where the headlines put it.

What it costs — the actual point of this release

Image output dropped from $60 to $30 per million image tokens, a 50 percent cut. Text input stays at $1.50 per million tokens, and text output with thinking is $7.50 per million. Here is what that becomes per image, on Google's published table:

Two things in that table are easy to miss and both matter more than the headline.

First, 4K fell by half but 1K fell by half too — and the gap to Pro widened. Pro's 1K price is 13.40¢, four times 2.1's 3.36¢, and Pro has no cheaper tier to fall back on. If you were rationing generations because of cost, that constraint has largely gone.

Second, batch mode halves it again. At $15 per million image tokens, a 1K image lands near 1.68 cents — roughly 4.6 times cheaper than the on-demand 1K price the old model charged. For anyone generating variants in volume, that is the number to build a pipeline around, and it is the single largest lever on this page.

Nano Banana 2.1 vs Nano Banana Pro — the honest split

This is the comparison everyone searches for, and it has two correct answers depending on whether you read the benchmark or look at the picture. Google's published numbers, on its own evaluation suite:

On paper it is not close. In practice, independent testing with a deliberately awkward prompt — a monkey holding a pink banana on a tiger, with a horse riding an astronaut in the background — found 2.1 followed the instruction, including the horse being above the astronaut, while Pro's version looked better: more natural colour, more convincing proportion. The horse at 2.1's scale read closer to a pony.

There is a caveat inside the caveat, and it is the reason to trust the benchmark a little less than the number suggests: Nano Banana 2, the outgoing model, reportedly matched Pro on those same tests too. When two consecutive generations both beat the flagship on the same suite and neither displaces it in real use, the suite is measuring something narrower than image quality.

What it still gets wrong

Google's model card is unusually direct about the first three, which is worth crediting — most launch pages would have omitted them.

• Small text still renders poorly. Larger typography and infographic layout improved markedly; fine print did not. Do not plan a poster around 8-point type.
• 3D spatial reasoning is limited. Depth, occlusion and how objects sit in a scene are the weak point, which is consistent with the scale error in the test above.
• World knowledge lags the backbone. It runs on Gemini 3.6 Flash, and its factual grounding trails that model rather than matching it. If the image depends on a specific real thing being correct, verify it.
• Relative scale — our addition, not Google's. The independent test's horse-shaped-like-a-pony problem is the practical form of the spatial limitation, and it is the one that gets a client comp rejected.

Where you can use it, and what the thinking levels buy

Nano Banana 2.1 is rolling out across the Gemini app, AI Mode in Google Search, Google AI Studio, Flow by Google, Stitch by Google, Google Ads and the Gemini Enterprise Platform. Developers can reach it in Google AI Studio and through the Gemini API.

Two capabilities are new enough to change how you prompt it.

Thinking levels. The model exposes minimal, medium and high — trading latency for quality. The difference is not cosmetic: infographic accuracy nearly doubles between the lowest and highest setting (0.328 to 0.521), and even text-to-image preference moves 35 points. Before you compare any quoted score with anything, check which level it was measured at, because a "2.1 beats Pro" claim is really a "2.1 at high thinking beats Pro" claim.

Reference capacity. Up to 14 reference images in one prompt, holding up to four characters and ten objects consistent through the generation. It also connects to Google Search for grounding. This is what makes the model credible for anything that repeats — a product carousel, a serialised character, one campaign with the same subject in many settings — and it is the capability the previous generation did not have.

The 29 October deadline

Google is shutting down gemini-3.1-flash-image — the API identifier for Nano Banana 2 — on 29 October 2026. Anything still pointing at that model breaks on that date.

The migration itself is trivial: change the model identifier to gemini-nano-banana-2.1. What is worth checking is everything downstream of it. Your per-image cost changes, your output resolution options expand to 1K, 2K and 4K, and the new thinking parameter did not exist on the old model, so any wrapper you built around it has a knob it does not know about. Teams that migrate by editing one string and moving on will get better images and an unexplained bill; the two minutes spent checking the resolution and thinking parameters first is the cheaper path.

Getting better images out of it

1. Draft at low thinking, finish at high.

   Thinking level is a cost and quality dial, not a default. Composing at minimal and re-rendering the composition you keep at high gives you most of the quality for a fraction of the spend — the same logic as drafting video at low resolution.

2. Use batch mode for anything counted in tens.

   At $15 per million image tokens it is roughly 4.6 times cheaper than the old on-demand 1K price. If your workflow is "generate twenty variants and pick one", this setting is the whole margin.

3. Give it fourteen references, not fourteen adjectives.

   The reference capacity is the reason to use this model over its predecessor. Supply the character, the product and the style as images and let the prompt describe only what is new. Four characters and ten objects can be held consistent — that is a budget to spend, not a limit to avoid.

4. Check proportion before you check beauty.

   The known failure mode is relative scale, not aesthetics. Zoom out and ask whether each subject is the size it should be next to the others before you look at anything else — it is the fastest way to catch the error that matters.

5. Route by what the image has to survive.

   Thumbnails, carousels and ad variants at 1K or 2K: 2.1. Anything at 4K where colour and proportion are the deliverable: Pro, and accept the 24-cent image. Run both for a week on real work and let the reject rate decide rather than the benchmark table.

6. Keep the model name in config.

   Between the 29 October retirement of gemini-3.1-flash-image and the announced arrival of Gemini 4 Argon, this family will move at least twice more this year. A one-line config change is a migration; a find-and-replace across a codebase is a project.

Frequently asked questions

What is Nano Banana 2.1?

Google's Flash-tier image generation and editing model, released 6 October 2026 as gemini-nano-banana-2.1. It is built on Gemini 3.6 Flash and replaces Nano Banana 2 (gemini-3.1-flash-image) from February 2026. Google calls it the more efficient counterpart to Nano Banana Pro, not a replacement for it.

How much does Nano Banana 2.1 cost?

Image output is $30 per million tokens, down from $60 — about half. Per image that is 3.36¢ at 1K, 5.04¢ at 2K and 7.56¢ at 4K, against 6.70¢, 10.10¢ and 15.10¢ for the outgoing model. Input text is $1.50 per million tokens; text output with thinking is $7.50 per million. Batch mode halves the image rate to $15 per million, landing near 1.68¢ per 1K image.

Is Nano Banana 2.1 better than Nano Banana Pro?

Higher on Google's benchmarks, not reliably better in practice. With thinking on it scores 1050 on text-to-image preference against Pro's 935, and 1106 on multi-character consistency against 1011. Independent side-by-side testing found Pro still better on colour and proportion while 2.1 misjudged scale. Note that Nano Banana 2 matched Pro on the same suite too — when two generations in a row beat the flagship on paper without displacing it, the suite is measuring something narrower than quality.

Is there a free way to use Nano Banana 2.1?

No free API tier — the first image is billed. Free-ish use lives in the consumer surfaces: the Gemini app, AI Mode in Google Search, Flow and Stitch, subject to those products' own limits. For volume work, batch mode at 1.68¢ per 1K image is the realistic answer rather than a free tier.

What is the difference between Nano Banana 2.1 and Nano Banana 2 Lite?

Different models at a shared launch price. Nano Banana 2 Lite is gemini-3.1-flash-lite-image, the fastest and cheapest member of the family, built for rapid iteration and A/B testing. 2.1 is the larger Flash-tier model with a 1M-token context, up to 14 reference images, 4K output and three thinking levels. Lite is for volume; 2.1 is for composition control.

How many reference images can Nano Banana 2.1 take?

Up to 14 in one prompt, holding up to four characters and ten objects consistent across the generation. That consistency is the release's headline capability and the reason it is credible for anything recurring — a carousel, a serialised character, a campaign with one subject in many settings.

What can Nano Banana 2.1 still not do well?

Google's model card lists weak small-text rendering, limited 3D spatial reasoning, and world knowledge that lags the Flash backbone. Independent testing adds relative scale: in a multi-subject test the model rendered one subject markedly too small where Pro did not. For anything where proportion carries the message, treat it as a draft tool.

What are the thinking levels in Nano Banana 2.1?

Minimal, medium and high, trading latency for quality. The gap is large — infographic accuracy moves from 0.328 to 0.521 and overall preference from 1015 to 1050. Treat high as a different cost-and-quality point rather than a toggle, and check which level any quoted benchmark used before comparing it.

Is Nano Banana 2.1 being deprecated or replaced soon?

Not 2.1, but the model it replaces has a deadline: gemini-3.1-flash-image shuts down on 29 October 2026 and anything still calling it breaks. Google has also announced Gemini 4 Argon as a forthcoming Pro-tier model, which is what would move the quality ceiling again.

Nano Banana 2.1 or Nano Banana Pro — which should you pick?

Pick on volume and on what the image has to carry. Pro costs about four times as much per 1K image (13.40¢ against 3.36¢) and offers no cheaper tier, but it remains safer for 4K hero work where colour and proportion are the deliverable. For social, product carousels and ad variants at 1K or 2K, 2.1 is the sensible default. The two are not the same width of tool either: 2.1 takes more references and has thinking levels; Pro's advantage is realism.

Also on this site

• Gemini Omni — Google's video model from the same family, and the one it built instead of a Veo 4. The free route, the real per-second cost, and the two leaderboards that disagree.
• GPT Image 2.5 — the direct competitor: two API models at one price, native transparent backgrounds, and a free tier this one does not have.
• Seedance 2.5 — ByteDance's video model, and the four specifications no engineering source confirms.
• MiniMax H3 — the video model that ranks first on Arena's image-to-video board, and the only one in this list you can download and run on your own hardware.
• Higgsfield AI review — the aggregator route: what each plan grants, what a clip really costs, and which surfaces bill even on an unlimited model.