Choose an image API model by requirements, not its name

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.

Five entries, one configuration checklist

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.

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
• 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

References

• API reference