OpenAI released ChatGPT Images 2.5 on September 8, claiming image generation latency up to 50% lower than Images 2.0 and better preservation of details through successive edits. The rollout covers ChatGPT, ChatGPT Work, and Codex across all tiers. Developers get two API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
The editing claim matters for a product image that needs a new background, then revised copy, then a different crop. A useful result must preserve the product through every change. If each revision damages something already approved, a faster generator can still leave the designer with more repair work. This release should be judged on how much of that work it removes. S5 Labs has not run a comparative image test; the performance claims here are OpenAI’s.
Flare and Sunburst make different tradeoffs
OpenAI positions Flare as the default for most applications, while Sunburst offers greater editing precision with longer generation times. The ChatGPT release also adds drawing through Sketch, image comments for focused edits, templates, and shareable prompts. Those interface features should help people communicate a brief, but they are separate from whether the model faithfully executes it.
The API guide gives developers a second decision. The Image API handles individual generation or editing requests. The Responses API supports conversational editing, retaining image context across turns. With Responses, the image model is selected inside the image-generation tool; the top-level model handles the conversation.
That distinction affects the application design. An editor needs to know which image a revision applies to and which version the user accepted. Keeping those versions visible would make it easier to recover from a bad change without asking the model to reconstruct earlier work. Even with better model consistency, the application still needs to track accepted versions.
The speed-versus-quality choice also connects to our coverage of Google’s Nano Banana 2 and Microsoft’s MAI image model launch. These are alternative products to evaluate against the same creative brief. Separate vendor demonstrations cannot establish which will finish that brief with the least intervention.
Measure a finished asset
A useful comparison would start with an approved source image and a fixed sequence of revisions. Ask both models for the same changes, then inspect the final images alongside the original. Record unintended changes to packaging, lettering, and any details the brief explicitly protects. Reviewers should score those defects before seeing which model produced each image.
Include review and repair time in the result. A model that returns quickly but needs several retries may cost more per accepted asset than a slower one. Conversely, extra generation time buys little when both outputs already meet the brief. Test at the intended delivery size: defects that disappear in a thumbnail can remain obvious in a presentation or product page.
OpenAI’s API documentation still lists limitations in text placement, recurring brand elements, and precise composition. It also warns that complex prompts can take up to two minutes. The headline latency improvement is therefore no guarantee that a particular editing session will finish quickly. Keep final lettering and layout checks in the acceptance process.
Provenance is part of the release
The Images 2.5 system card says OpenAI combines C2PA metadata with Google DeepMind’s SynthID invisible watermarking. These mechanisms help identify generated content; they do not verify the factual accuracy of what an image depicts.
The card also reports safety tests using adversarial prompts and explicitly says those prompts do not represent ordinary production traffic. Its lower aggregate unsafe-output figures should not be read as a measured incident rate for a creative application. OpenAI notes that none of the unsafe-shown differences reaches its stated statistical-significance threshold.
An editorial team still needs to distinguish an invented scene from documentary evidence, regardless of its provenance marker. For product imagery, an attractive result still needs comparison with the actual product. For both uses, the test should end with an asset someone can approve without overlooking an altered detail.
Sources
- Introducing ChatGPT Images 2.5 — OpenAI
- Image generation — OpenAI API documentation
- ChatGPT Images 2.5 System Card — OpenAI