Nano Banana 2.1 halves what Google charges for a 1K or 2K image and triples what it charges for everything else. At 4K it spends 3,780 tokens, so the cut there is 25%.

If you only have a minute

• Images cost half what Nano Banana 2 charged at 1K and 2K ($0.0336, $0.0504). At 4K you save a quarter, and the real price is $0.113, not the $0.0756 some coverage printed.
• Prompt tokens now cost $1.50 a million and thinking $7.50. A normal request stays well ahead; a 1K image only loses its discount past ~33,600 input tokens or ~7,470 thinking tokens.
• Nano Banana 2 is deprecated with no shutdown date. Lite now costs the same per 1K image and scores lower on LMArena.

The rate card, line by line

Nano Banana 2.1 went GA on the Gemini API on October 6 as gemini-nano-banana-2.1. Google dropped the gemini-3.x-flash-image naming with it. Here is every per-million line next to the model it replaces.

Gemini API pricing page, read October 12, 2026. No free tier for either model. Batch halves every line on both.

Input and thinking got more expensive. Image output got cheaper, and image output is usually most of the bill. A 1K image is 1,120 output tokens; a typical prompt is a few hundred.

Per image: half price, except at 4K

Google's printed per-image prices, image output only. Lite stops at 1K. Token counts for 2.1: 1,120 / 1,680 / 3,780.

At 1K and 2K the token counts didn't move, so the halved rate halves the image. At 4K, 2.1 spends 3,780 tokens where Nano Banana 2 spent 2,520, 50% more, and the saving shrinks to 25%: $0.151 down to $0.113.

That's also why you'll see $0.0756 quoted for a 4K image in some launch coverage. It's 2,520 tokens times the new $30 rate, the old token count on the new card. Google's own footnote says 3,780, and 3,780 × $30 per million is $0.1134. If you budgeted 4K work off the lower number, the real cost is 50% higher.

How much prompt or thinking erases the saving

Each 2.1 request pays $1.00 per million more for input than Nano Banana 2 and $4.50 per million more for thinking. Thinking is on by default at medium, and you can lower it to minimal but not turn it off. So the question is simple arithmetic: how many extra tokens does it take to eat the image discount?

Our arithmetic. Savings use exact token counts (1,120 × $60 per million is $0.0672, which Google rounds to $0.067). A mix of input and thinking breaks even proportionally.

A plain text-to-image call doesn't get close. With a 300-token prompt and 1,000 thinking tokens, a 1K image costs $0.0704 on Nano Banana 2 and $0.0416 on 2.1. At 3,000 thinking tokens it's $0.0764 against $0.0566, still 26% cheaper.

Editing is where to look. The image-generation docs allow up to 10 object references and 4 character references. Those 14 reference images would have to average 2,400 tokens each to wipe out the 1K saving on input alone, and that's before any thinking. Google doesn't publish a typical thinking count for either model, so check usage_metadata on a few real requests before you assume the halved number.

10,000 images a month

Image output only, standard tier. Nano Banana 2.1 in accent.

Google rates from the Gemini API pricing page; GPT Image 2.5 at 1024x1024 from our September breakdown. Batch halves every Google row.

GPT Image 2.5 at medium is still the cheapest thing on the chart by a wide margin, and it leads LMArena's text-to-image board (Sunburst at 1425). The arena doesn't say which quality setting it ran, so don't read that score as the $0.013 image's.

Inside Google's lineup the picture is cleaner. 2.1 sits 6th on that board at 1328 on an early 5,310 votes. Nano Banana 2 is 11th at 1262 and Nano Banana Pro 16th at 1248, for four times 2.1's price at 1K. Google's own launch Elo numbers circulated in press coverage, but we couldn't find them on any Google page.

Lite now costs the same per image

This is the part we didn't expect. Nano Banana 2 Lite also bills images at $30 per million and also spends 1,120 tokens on a 1K image. So a 1K image is $0.0336 on both, to the cent. On LMArena, 2.1 scores 1328 to Lite's 1250.

Lite keeps two advantages. Its input is $0.25 per million and its thinking $1.50, a sixth and a fifth of 2.1's. At a 300-token prompt and 1,000 thinking tokens that's $0.0352 per image on Lite against $0.0416 on 2.1, about 18% more for 2.1. If you generate at 1K with short prompts and quality is good enough on Lite today, there's no rush. If you've been tolerating Lite for the price, 2.1 is roughly that price.

Nano Banana 2 is deprecated, with no date

The October 6 changelog entry deprecates gemini-3.1-flash-image and says to move to gemini-nano-banana-2.1. No shutdown date yet, and its prices haven't changed. Given how Gemini 2.5's dates moved around, we wouldn't wait for one.

For most callers this is the rare migration that lowers the bill. Swap the model ID, log input and thinking tokens for a day, and compare against the break-even table above. You can run your own mix through the calculator, or see every rate on the Nano Banana 2.1 model page.

Google's docs, the leaderboard, and the launch story

• Gemini API: pricing - every rate and per-image figure for 2.1, Nano Banana 2, Lite and Pro
• Gemini API: changelog - October 6 GA and Nano Banana 2 deprecation
• Gemini API: image generation - thinking levels, resolutions, reference image limits
• Gemini API: models - model ID and stable status
• LMArena: text-to-image leaderboard - scores as of October 7, 2026
• Decrypt: Google launches Nano Banana 2.1 - launch coverage, Google's internal Elo figures