Nano Banana 2 Billing: 1,120 Tokens Flat, and the Pro Surcharge Nobody Documents

Nano Banana 2 bills exactly 1,120 tokens for a 1K image, and that number does not move no matter what you prompt. Its sibling Nano Banana Pro does move — it adds reasoning tokens that scale with prompt length, on top of the same fixed image tokens. That difference is the source of a long-running billing confusion, and it is not in either model’s documentation.

Everything below was measured against the live endpoint on 7 September 2026.

Nano Banana 2:      google/gemini-3.1-flash-image
Nano Banana 2 Lite: google/gemini-3.1-flash-lite-image
Nano Banana Pro:    google/gemini-3-pro-image
Endpoints:          /v1/images/generations, /v1/chat/completions
Image tokens:       1,120 at 1K · 1,680 at 2K · 2,520 at 4K
Reasoning tokens:   none on 2 and Lite · 223–377 measured on Pro
Rate:               $60 per 1M output image tokens (Nano Banana 2)

The Measurement

Two prompts against google/gemini-3.1-flash-image — one of eight tokens, one of fifty-one — produced identical billing:

The output side is a constant. Prompt length changes what you pay on input, which at $0.50 per million is rounding error, and nothing else.

Now the same test on google/gemini-3-pro-image:

Pro’s image half is fixed and its reasoning half is not. A longer prompt bought 69% more reasoning tokens. This is why a per-image price computed from the token table alone always understates Pro, and never understates Nano Banana 2.

If you have been trying to reconcile an image bill against a quoted per-image figure, that is the reconciliation: check output_tokens_details and see which of the two fields is present.

{"output_tokens": 1343,
 "output_tokens_details": {"image_tokens": 1120, "reasoning_tokens": 223},
 "total_tokens": 1357}

The 512px Trap

Asking for the cheapest tier does not get you the cheapest tier. We passed size: "512x512", decoded the returned PNG, and measured the actual pixels:

The three larger sizes are honoured exactly and bill exactly as Google’s token table says. The 512 request is silently upgraded. Google’s table prices a 0.5K image at 747 tokens; you get 1,120 instead, so a draft you budgeted at $0.0448 costs $0.0672 — 1.5x.

The lever that does work is the model. google/gemini-3.1-flash-lite-image returned the same 1,120 tokens at half the rate, $0.0336 per image. For draft-and-select workflows, switch model rather than shrinking size.

What It Actually Costs

Computed from the measured token counts at the live catalog rates:

Pro is quoted as approximate on purpose: $0.1344 of it is fixed image tokens and the rest is reasoning, which moved between $0.0027 and $0.0045 across our two prompts. It is the only model of the three whose per-image cost you cannot state as a single number.

These match Google’s own list at every tier — 1,120 × $60/M is $0.0672, and so on. The rate card is not where the surprises are; output_tokens_details is.

Calling It

curl -X POST https://api.ofox.ai/v1/images/generations \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.1-flash-image",
    "prompt": "A single red apple on a white table",
    "size": "1024x1024"
  }'

Returns data[0].b64_json. Read usage.output_tokens_details on the response and you will know your real per-image cost immediately, rather than inferring it from a pricing page.

Use /v1/images/generations, not /v1/chat/completions. The chat route accepts these models and returns HTTP 200 with finish_reason: "stop" — and no image. We measured content of length zero with no images key on the message, while still being billed the full 1,120 tokens. It looks like a successful call and costs like one.

The Error Bodies

Captured live, so you can match them exactly:

Marketing name instead of model ID. “Nano Banana 2” is not a callable string:

{"error":{"message":"Model 'google/nano-banana-2' not found","type":"model_not_found","code":404}}

Missing prompt:

{"error":{"code":null,"message":"prompt cannot be empty","param":null,"type":"invalid_request_error"}}

Batching. n: 2 fails with an upstream Google error that leaks through the gateway:

{"error":{"message":"Invalid JSON payload received. Unknown name \"numberOfImages\" at 'generation_config.image_config': Cannot find field.","type":"INVALID_ARGUMENT"}}

There is no batch parameter on this route. Loop over single requests.

A size the model will not take returns a generic api_error rather than a validation message, so if you get “An error occurred while processing your request” on an image call, suspect your size string before anything else.

What to Take Away

• On Nano Banana 2, prompt length is free on the output side. Write the long prompt. It costs the same 1,120 tokens as the short one.
• On Nano Banana Pro, prompt length is not free. Budget it as a range, not a number, and read reasoning_tokens before scaling.
• Do not use size as a cost lever below 1K. Use Lite.
• Alarm on usage, not on the rate card. Both facts above are visible in output_tokens_details on every response and invisible in any pricing table.

Related

• GPT Image 2 release guide — the other flagship image model, with a different billing shape again.
• GPT Image 2 generation failures — error handling on the OpenAI image route.
• Qwen 3.8 Flash reasoning tokens — the same reasoning-token billing trap on the language side, where a low ceiling returns empty content you still pay for.

Sources

• https://ofox.ai/models/google/gemini-3.1-flash-image
• https://ai.google.dev/gemini-api/docs/pricing
• https://discuss.ai.google.dev/t/i-am-confused-on-pricing-for-nano-banana-pro-i-am-getting-2000-output-tokens-instead-of-1120-for-a-1k-image/121299

Model IDs, rates and endpoints were read from the live Ofox /v1/models endpoint on 7 September 2026. Token counts are from live /v1/images/generations and /v1/chat/completions calls the same day: four sizes and two prompt lengths against google/gemini-3.1-flash-image, two prompt lengths against google/gemini-3-pro-image, and one call each against google/gemini-3.1-flash-lite-image. Output dimensions were verified by decoding the returned PNG and reading the IHDR header rather than trusting the request. Error bodies are quoted verbatim from live responses. This is a small sample on one route on one day, not a specification — re-run it on your own account before building a cost model on it, and expect undocumented behaviour like the 512px upgrade to change without notice. The linked Google forum thread reports a related discrepancy on Google’s direct API that was still open when last updated in March 2026; we did not reproduce that thread’s numbers and are not claiming its cause.