Nano Banana comparison: five models, real images and prices

Which Nano Banana name maps to which model?

The table below maps all five Nano Banana names to exact MaxAPI request IDs. They are separate image entries: Nano Banana 2.1 is not an alias for Nano Banana 2, and Nano Banana 2 Lite is not the text-only Gemini 3.1 Flash Lite. Choose by reference count and output controls before comparing cost. Nano Banana 2.1 offers 1K, 2K and 4K; do not copy Nano Banana 2’s 512 tier into a 2.1 request.

Start with the constraints that can disqualify a model

Write down the number of references, required aspect ratio, target output size, whether your application needs explicit quality controls, and which routes your account can use. A model name containing Pro or Flash is not proof of better output or lower latency on your workload. Do not choose a text-only Gemini entry for image generation. Keep preview, beta and platform aliases distinct in your saved configuration.

Exact model IDs and configuration

Reference counts are configured maxima, not a guarantee that every file or route will accept every input. Resolution labels are the choices exposed by this frontend. The size sent on the wire is not interchangeable across families: Gemini uses resolution and aspect_ratio, while GPT Image uses a concrete size. Publicly listed and permission-gated routes are shown separately; neither list proves your key currently has access.

What changed the model choice in this test?

All 21 submitted jobs returned an image, at a total settled cost of 0.835 credits. All fifteen 1K outputs were 1024 × 1024. The six additional Pro/2/2.1 outputs matched the requested 2048 × 2048 and 4096 × 4096 sizes. This describes this batch only, not a long-run success rate.

For text-heavy posters, Lite and Pro met the requested wording and flat layout in these samples. Original Nano Banana misspelled the English headline and the final Chinese character; 2 and 2.1 added a paper mockup that the prompt prohibited. All five edits added the gray background and orange sphere, but Pro noticeably reduced the mug size. Inspect the originals for your own acceptance criteria.

PNG was requested for every job, but four returned files were JPEG: 2 product, Pro product, Lite edit and 2 at 4K. Check the file bytes rather than assuming output_format was honored. Original files below keep their actual format.

What do 1, 100 or 1,000 Nano Banana images cost?

Prices below are the Mix quote checked on October 16, 2026, with n=1 and no account discount. 1 credit equals US$1. Original Nano Banana and 2 Lite bill per call; Pro, 2 and 2.1 bill per output image on this route. For the listed 1K/2K/4K labels, this quote has no separate resolution surcharge. A supported label does not prove the delivered pixel size; inspect the tests below.

The 100/1,000 budgets multiply the single-call price and assume one completed image per call. They exclude retries, taxes, discounts, optional tools and application costs. Official token-priced routes need actual returned usage and must not use this flat-price table. Check /v1/model-pricing with your own key and remember that a key-pinned route overrides a payload route.

How we tested the five models

We preselected three tasks: a product photo, an English/Chinese poster and an edit of one shared mug reference. Each model received the same prompt for that task, n=1, 1K, 1:1, PNG requested, URL response, async=true and a key pinned to Mix. We did not set a token cap, thinking level, search tool, seed or GPT-style quality field. We kept the first completed output without aesthetic rerolls.

Product checks: one blue mug, right-side handle, full object, ivory background and no extra objects. Poster checks: all three requested lines readable with exact wording, no extra text, blue mug below and orange circle at upper right. Edit checks: gray background, one orange sphere on the left, recognizable glaze and handle, and any change in mug scale or placement. These are small qualitative observations, not model-wide success rates or a statistical ranking.

Times run from client submission to observed completion and include network transfer and polling. Two jobs could be in flight. Route selection and fallback may differ between runs; model names do not independently verify upstream weights. This is not a speed benchmark. Costs come from settled usage records, not reservations.

Results, output dimensions and actual charges

Product photography: same prompt

A studio product photograph of a single cobalt-blue ceramic mug on a warm ivory background. The handle is on the right. Show the entire mug, centered with generous empty space, soft lighting and a subtle shadow. Square composition. No text, logos, watermark or other objects.

Typography: Chinese and English

Design a flat square poster for a fictional coffee shop. Warm ivory background, navy typography, one blue ceramic mug in the lower half and one small orange circle in the upper right. Include exactly these three text lines, each on one line, with correct spelling and no other text: "SLOW MORNINGS", "咖啡与好时光", "SATURDAY 9 AM". Keep all letters inside generous margins. Do not show a frame, wall or paper mockup.

Editing: one shared reference

Edit the reference image. Preserve the same blue ceramic mug, its shape, glaze texture, proportions, position and right-side handle. Replace only the ivory background with a pale gray studio background and place one small orange sphere to the left of the mug. Keep the entire mug visible. No text, logos or other objects.

Does the resolution label match the delivered pixels?

1K product images above and the 2K/4K runs below use the same prompt. Each is a separate generation, not an upscale or a controlled same-seed pair. Inspect actual pixel dimensions and file format.

Reproduce the request and inspect the evidence

The evidence download contains the exact prompts and request settings, observed status, settled credits, output dimensions and SHA-256 hashes. Originals are self-hosted without retouching; small previews are for page loading. The test is dated evidence, not a live availability guarantee. Use the migration guides for a runnable reference-edit example and the model pages for complete API parameters.

Download prompts, settings, charges and image hashes · MaxAPI / fal.ai · MaxAPI / Replicate

Apply the table to an actual integration

For a first 1K workflow with no more than three references, the two 1K-only entries satisfy those catalogue constraints; this is not a claim that either is cheapest or fastest. If you require fourteen references or selectable 2K/4K output, those 1K-only entries fail the requirement before quality testing starts. The listed Nano Banana Pro and Nano Banana 2 entries have larger configured reference limits. GPT Image 2 exposes named quality settings; do not copy those settings into a Gemini request without checking its contract.

A 4K label does not mean a 4096 × 4096 image. For GPT Image 2, the current square 4K option sends 2880x2880, and the 16:9 option sends 3840x2160. For Gemini, send its documented resolution label and ratio and inspect the returned image dimensions. Confirm the actual bytes and dimensions before accepting a result into your application.

Compare the same billable job, not headline prices

Open each model’s price section and your signed-in live pricing page. Compare the route, output count, size, quality and billing unit together. Record a quote date; a catalogue price is not an account-specific final quote. Key-pinned routes can override a route requested in your payload. Include failed or repeated jobs and refunds in your own cost review without assuming every failure is charged or refunded the same way. Credits, cash paid and business margin are different measures.

Use a small, reproducible acceptance test

Use only owned or licensed, non-sensitive references and the same task description for each candidate. Save the public model ID, effective route, exact payload, input order, request ID, completion time, decoded output size and settled cost privately. Evaluate the qualities your product needs—such as subject preservation, readable text or background consistency—using a fixed rubric, not one attractive sample. Set a spending cap before testing and report both successes and failures. This comparison deliberately supplies no invented benchmark scores.

Model-specific parameters

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

• Gemini 2.5 Flash Image google/gemini-2.5-flash-image
• Gemini 3.1 Flash Lite Image google/gemini-3.1-flash-lite-image
• Gemini 3.1 Flash Image Preview google/gemini-3.1-flash-image-preview
• Nanobanana Pro google/gemini-3-pro-image-preview
• Nano Banana 2.1 google/gemini-nano-banana-2.1
• GPT Image 2 openai/gpt-image-2

Related guides

• Image resolution, aspect ratio and quality are different controls
• Send reference images without avoidable input errors
• Understand routes, prices and available balance
• GPT Image 2 vs Nano Banana 2 vs Pro: six real outputs

References

• API reference