GPT Image 2 vs Nano Banana 2 vs Pro: six real outputs

What we tested

We submitted one request per prompt and model through the MaxAPI website. One earlier GPT Image 2 product sample was reused; the other five images came from five new user submissions. No manual rerolls, retouching or selection of winners. Gemini used resolution=1K and aspect_ratio=1:1; GPT used size=1024x1024 and quality=high. All used route_mode=mix and n=1, without reference images.

The GPT request asked for high quality, but the task response did not expose the effective quality tier. This is not independent confirmation of upstream high quality.

Charges are this test account’s settled credits on the test date, not permanent prices. Timings are individual observations, not a speed ranking or latency promise.

Product photo · blue ceramic mug

The same English prompt was sent unchanged for all three outputs. Compare subject scale, position, surface texture and empty space; these observations describe only these images.

A single blue ceramic mug on a plain white background, soft studio lighting, centered square composition, no text, no logos.

Text layout · coffee-shop poster

The prompt requests three exact English strings and exactly three lines. Check the actual line breaks as well as spelling. The prompt remains in English in every language version of this article because translating it would change the test input.

The three posters contain the requested words, but each wraps the headline into two lines. Correct spelling did not satisfy the strict three-line layout requirement.

Create a square editorial poster for a fictional independent coffee shop. Warm ivory background, a blue ceramic coffee cup in the lower half, a single orange circle in the upper right, restrained navy and orange palette, generous margins, crisp typography. Include exactly these three lines of English text and no other text: "SLOW MORNINGS" as the large headline, "COFFEE & CONVERSATION" as the subtitle, and "SATURDAY 9 AM" as the small footer. Keep every letter readable and all text inside the canvas. No logos, no watermark.

Completion, recovery and cost

All six images completed at 1024 × 1024. The two Nano Banana Pro submissions each encountered an initial upstream HTTP 503 and completed after Mix automatically selected another channel. The five new user submissions therefore produced seven upstream attempts. A recovered attempt is not another user request, and should not be hidden when reviewing the route.

Service completion is the recorded request duration. Browser result ready includes the frontend submission and result-check workflow. GPT returned an asynchronous task and was polled; Gemini returned results synchronously. Different channels and execution paths mean these timings are not a controlled model speed comparison.

Per-image settled charges were 0.019 credits for GPT Image 2, 0.045 for Nano Banana 2 and 0.065 for Nano Banana Pro. The six outputs totalled 0.258 credits, including the reused 0.019-credit sample. The five new submissions cost 0.239 credits. These are observed settlements, not a quote for another account or date.

Use the results without overgeneralizing

Two samples per model are not a reliability study. Mix can select different upstream channels, so these outputs do not independently certify provider identity or quality. Reference editing, 4K, batch consistency and other routes were not tested.

All three posters spelled the requested words correctly, yet none kept the strict three-line layout. For publication-ready designs, review spelling, line breaks, margins and object placement separately. A successful HTTP response is not visual approval. If exact typography is essential, consider adding the final text in your design tool.

Start with a small test using your own non-sensitive task, inspect current account pricing and set a spending cap before submitting. Record both successes and failures. Six completed examples cannot establish a long-term success rate or guarantee commercial suitability.

Model-specific parameters

Public configuration snapshot; not live availability. Follow each model page for its parameters, supported routes and reference limits.

• GPT Image 2 openai/gpt-image-2
• Gemini 3.1 Flash Image Preview google/gemini-3.1-flash-image-preview
• Nanobanana Pro google/gemini-3-pro-image-preview

Related guides

• Choose an image API model by requirements, not its name
• Image resolution, aspect ratio and quality are different controls
• Understand routes, prices and available balance
• Diagnose image API errors before resubmitting

References

• API reference