On September 8, 2026, OpenAI released ChatGPT Images 2.5 across ChatGPT, ChatGPT Work, Codex, and its API. The update is less about a single headline benchmark than a workflow change: faster generation, more selective edits, stronger reference-image fidelity, and two API models aimed at different production needs.
OpenAIWhat changed in ChatGPT Images 2.5
OpenAI says the model preserves recognizable subjects more reliably when a reference photo is transformed, keeps lighting and texture more natural, and changes only the requested part of an image more consistently. The multi-turn claim matters most in practice. Image editing often fails not on the first instruction but on the fourth, when an earlier face, layout, or brand detail drifts while something else is changed.
The ChatGPT interface also adds Sketch, which accepts a rough drawing as a spatial reference, templates for formats such as posters and product photos, comments placed directly on an image, and prompt sharing. These tools narrow the gap between describing an edit and pointing at it.
OpenAI claims up to 50% lower generation latency
Relative latency index, with Images 2.0 set to 100. This is a vendor claim, not an independent benchmark.
Source: OpenAI, Introducing ChatGPT Images 2.5, September 8, 2026.
The speed figure needs careful wording. OpenAI reports latency reduced by up to 50% compared with Images 2.0, while an early Axios hands-on test described the editor as faster and better at preserving people and pets. Neither is a controlled third-party benchmark across a representative prompt set. The supported conclusion is that speed and edit stability improved, not that every generation will take half as long.
Flare versus Sunburst in the API
Developers get two named models. GPT-Image-2.5 Flare is the default option for rapid iteration and high-volume use. GPT-Image-2.5 Sunburst is positioned for detailed creative work where edit precision matters more than generation time. Both can be selected through the Images API or the Responses API image-generation tool.
One release, two API operating points
Both models use the same published token rates; the tradeoff is speed versus editing precision.
Faster iteration and high-volume creative workflows
Longer generation for detailed editing and polished assets
Source: OpenAI model documentation and pricing, accessed September 9, 2026.
The published token rates are the same for both: $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million image-output tokens. Cached inputs are cheaper. OpenAI says those token rates match GPT Image 2, although the older calculator does not estimate 2.5 token consumption. That caveat matters because identical per-token rates do not guarantee identical cost per finished asset.
What the release means for creative workflows
The most consequential capability is selective persistence: keeping the parts a user did not ask to change. That makes image models more useful in production settings where a team iterates on one approved composition, product, or character rather than generating disconnected concepts from scratch. OpenAI specifically points to on-brand asset series, UI concepts, visual search, product imagery, and campaign creative.
Distribution is equally important. Adobe says the models are already available through Firefly, while OpenAI names Higgsfield, Manus, and Runway as early users. Image generation is becoming an embedded capability inside creative systems rather than a destination product on its own.
Safety remains a declared rather than independently audited part of the release. OpenAI says Images 2.5 uses prompt and image checks, C2PA metadata, and invisible watermarking. Those controls help with provenance, but they do not by themselves prove that every downstream platform preserves the metadata or that every synthetic image will be recognized.
For users, the practical test is not whether a launch gallery looks better. It is whether a subject survives repeated edits, whether text and layouts remain usable, and whether the cost of reaching an approved result falls. Flare and Sunburst make that tradeoff explicit: choose speed for exploration, precision for the final pass, and measure the full revision loop rather than a single render.
Sources:
- Introducing ChatGPT Images 2.5 | OpenAI
- GPT-Image-2.5 Sunburst model and pricing | OpenAI API
- Hands on with ChatGPT's new image editor | Axios
Image credits
Header and in-body photograph: 1515 Third Street in San Francisco, photographed by Coolcaesar via Wikimedia Commons, licensed under CC BY 4.0. The image was visually reviewed before use and does not depict the product release.