An image model can make a convincing first draft and still miss the detail that matters: the label on a product, the shape of a garment, or a layout that has to stay fixed. That is the useful starting point for OpenAI’s latest image release.
OpenAI’s API changelog records GPT Image 2.5 Sunburst and Flare as released on September 8. The company positions Sunburst for work where editing precision matters most, and Flare for fast, high-quality everyday generation. Those are vendor descriptions. The right choice for a real product still needs a small, repeatable evaluation.
Two models with different starting points
Both models accept text and image inputs and support image generation and editing. They can be selected through the Image API or the image-generation tool in the Responses API. The release also adds xhigh and max quality settings alongside the existing choices.
The model names give teams a practical way to start a comparison. A workflow with demanding edit constraints may start with Sunburst. A workflow that already produces acceptable results may start by checking whether Flare can meet the same requirements with less waiting. Access can depend on account requirements, including API Organization Verification.
Keep the test close to the work
For a product catalog, the test set should include the difficult images: reflective packaging, small lettering, unusual proportions and similar colors that are easy to confuse. For a design tool, include the edits users actually request after the first result.
Keep the prompt, reference files, output dimensions and explicitly selected quality setting consistent for the first comparison. Record what passes the review and what needs another generation or manual correction. A pleasing image is only useful if it still represents the intended product and fits the destination.
- Check identity and proportions before judging style.
- Inspect text, brand colors, edges and any required transparent background.
- Measure time and cost until a reviewer accepts the result, including retries.
Compare the cost of a usable result
OpenAI’s documentation cautions that matching token rates do not imply matching cost per image. Token use varies with the model, quality, dimensions and prompt. A Responses API workflow also includes the mainline model’s token usage.
This matters when an image feature moves from a demonstration into a product. The useful number is the cost of an accepted asset. A cheaper attempt can become expensive if it produces more retries, and a higher quality setting is worthwhile only when it solves a requirement that is still failing.
Sources and release notes: [4]
A sensible next step
Our recommendation is to choose one existing workflow and define its acceptance criteria before replacing the model behind it. Keep a small set of approved inputs and compare both options against that same set.
We have not benchmarked these releases for this article. The documentation still lists limitations around precise text, consistency and structured composition. A short evaluation with real requirements will tell a team more than a broad promise about image quality.
Sources and release notes: [4]
Sources and release notes
Original analysis by AtmosGate, based on the sources below. Vendor claims and release status are attributed to their publishers.
- API changelog: September 8, 2026
OpenAI · 8 September 2026
- GPT Image 2.5 Sunburst model documentation
OpenAI
- GPT Image 2.5 Flare model documentation
OpenAI
- Image generation guide
OpenAI