Nano Banana 2 (Gemini 3.1 Flash Image): Resolution, Cost, Fit

The most common mistake with Gemini 3.1 Flash Image is budgeting it like a text model. Its catalogue price, around $0.50 in and $3.00 out per million tokens, describes text and thinking tokens. The images are billed at a separate, much higher rate, and how much each one costs depends almost entirely on the resolution you ask for.

Get that right and Nano Banana 2 is easy to reason about. It is Google's middle image model: more capable than Flash Lite Image, much cheaper than Gemini 3 Pro Image, and built for work where 1K is not always enough.

A short timeline, because it matters

The model has changed status three times in eight months, and which endpoint you call depends on when you integrated:

• 26 February 2026 — launched as gemini-3.1-flash-image-preview, which Google described as optimised for speed and high-volume use.
• 28 May 2026 — the stable gemini-3.1-flash-image reached general availability. The preview was deprecated the same day and shut down on 25 June 2026.
• 6 October 2026 — Google released Gemini Nano Banana 2.1 (gemini-nano-banana-2.1) as an update to this model and deprecated gemini-3.1-flash-image. Google's changelog says no shutdown date has been announced.

Two practical consequences. If your code still names the -preview model, it has been failing since June. And if you are starting fresh, evaluate 2.1 alongside this one rather than building on a model already marked for retirement. Pinning model versions explains why the stable ID, not a moving alias, is what you want in production either way.

Resolution decides the bill

Google's pricing page lists image output at $60 per million tokens, with a fixed token count per resolution. In per-image terms on the standard tier:

• 0.5K — $0.045
• 1K (the default) — $0.067
• 2K — $0.101
• 4K — $0.151

The Batch API roughly halves each figure, to $0.034 at 1K and $0.076 at 4K. Input is $0.50 per million tokens for text and images, and text or thinking output is $3.00 per million.

The useful pattern is to generate small and upscale only what survives. If a user picks one of four drafts, rendering all four at 4K costs $0.60; rendering four at 0.5K and the chosen one at 4K costs $0.33, and the user waits less for the drafts. The Batch API is worth considering for non-interactive renders that can wait.

What it does that the others do not

Google's model page lists a few capabilities that set it apart within the family:

• Image Search Grounding — the model can pull in web image results to render real subjects more accurately. Google's image generation guide lists it only for this model and its 2.1 successor. Search grounding is billed separately once you pass Google's free monthly allowance.
• Extreme aspect ratios — 1:4, 4:1, 1:8 and 8:1, useful for banners, vertical strips and panoramas.
• 4K output, which Flash Lite Image does not offer at all.
• Improved multilingual text rendering and aspect-ratio adherence, by Google's description.

It also accepts text, images, video and PDF as input, with a 131,072-token input limit and 32,768-token output limit. Thinking is always on: Google's image generation guide says it cannot be disabled in the API, though you can choose minimal or high. Function calling, caching and structured outputs are not supported, whatever some catalogue listings say, so do not design a pipeline that needs the model to return a JSON object alongside its image.

Choosing between Lite, Flash and Pro

The three tiers split along a clear line:

• Flash Lite Image — 1K only, around $0.034 per image, latency-first. Right for stickers, thumbnails, quick edits and anything interactive.
• Flash Image — 0.5K to 4K, $0.045 to $0.151, with search grounding. Right for marketing assets, infographics, diagrams with real subjects and anything that will be printed or shown large.
• Pro Image — $0.134 at 1K or 2K and $0.24 at 4K on Google's pricing page. Right when a single image is valuable enough to justify the premium and the work is complex.

At 1K, Flash costs about twice Lite. That premium is worth it only if the output is visibly better for your prompts, or if you need grounding. If neither applies, Lite is the better default and Flash is the escalation path, a pattern model routing and fallbacks describes in general terms.

Where Nano Banana 2.1 fits

Google describes 2.1 as an update to Nano Banana 2 with improved visual quality, multi-turn character consistency and text rendering, across 1K, 2K and 4K. Its pricing page lists 2.1 image output at $30 per million tokens on the standard tier, half the $60 rate for this model. 2.1 drops the 0.5K option, but if your resolutions are 1K and up it is both newer and cheaper, which leaves little reason to start a new integration on 3.1 Flash Image.

The reason to stay put is a tuned prompt library and a deadline. Then keep the stable ID pinned, rerun your prompts on 2.1, and switch when the outputs pass review.

How to test it on your own work

1. Collect thirty prompts from real usage, including text-heavy ones and any that depend on real-world subjects.
2. Run each at 1K on Lite, Flash and 2.1. Note cost and latency per image.
3. Rate blind on three things: text legibility, prompt adherence, and whether a designer would use it without edits.
4. For prompts where Flash wins, rerun with search grounding on and off to see whether grounding is the reason.