OpenAI released GPT Image 2.5 as two models, Flare and Sunburst

The New Stack reports OpenAI launched GPT Image 2.5 this week as a pair of models, Flare and Sunburst. Both are available on the API at the same posted token rates: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. The New Stack also says OpenAI claims Flare produces higher-quality images than GPT-Image-2 while running 50 percent faster.

Two models, one price sheet

Flare is described in the release as the faster default option for most applications, and Sunburst as the more precise choice for edits, with a longer generation time. The New Stack reports the API pricing table lists the same per-million rates for both, and OpenAI does not publish a per-image cost or an average token consumption for either model. So a team choosing between Flare and Sunburst for a repeatable job cannot read the cost difference off the pricing page; it has to be measured against actual traffic.

The New Stack quotes OpenAI positioning Sunburst as "built for premium visual workflows that benefit from tighter control across edits", and Flare as the model to choose when latency matters more than the last percent of fidelity. The article also notes that OpenAI has not said how Sunburst's precision translates to added token cost or added time compared to Flare.

What actually changes for image API users

The visible change in the release, per The New Stack, is a precision-editing capability that lets a developer change one region of an image (a product, a background, a piece of on-image copy) while the rest of the picture stays the same across multiple rounds of edits. That matters for teams running variant workflows, where each edit currently regenerates parts of the image that were already approved and then need re-fixing.

The second claim The New Stack quotes is greater image fidelity, so reference-led workflows keep a variation more closely tied to the source. Teams that build product-catalogue tooling, ad-variant pipelines, or asset localisation are the audience for that: the pattern of "here is the main photo, keep everything but the label" is where GPT-Image-2 has been losing details on every generation.

The unanswered question is the bill

Because OpenAI has not published a way to estimate token consumption per image, The New Stack points out there is no way to read the pricing sheet and predict what a given image will cost. Two models at the same per-million rate can still produce different bills for the same job when one uses more tokens per image. That is the missing piece: OpenAI has said the models cost the same per token, and has not said anything about tokens per image.

For a team already committed to GPT-Image for production imagery, the next step is to run both Flare and Sunburst against the same 20 or 50 reference images and record the token counts in the API responses before picking one. For a team on GPT-Image-2 today, Flare is the direct replacement with the 50 percent latency claim to check, and Sunburst is the upgrade to try only where the current output has been moving away from the source on repeat edits.

Source

• GPT Images 2.5 promises edits that leave the rest of your image alone, The New Stack, 10 September 2026.