Updated September 10, 2026. Rollout began September 8 on ChatGPT, ChatGPT Work, and Codex.
ChatGPT Images 2.5 is OpenAI’s new image generation and editing system. The update focuses on sharper details, natural lighting, subject retention in reference photos, and more reliable localized editing. OpenAI also claims up to 50% lower latency than Images 2.0. The template is available on desktop, mobile and web for all ChatGPT levels, as well as in Codex and ChatGPT Work.
What changes with ChatGPT Images 2.5
The most useful improvement is not a new default aesthetic, but the ability to change an element without rebuilding the entire image. In a product retouching, for example, it should be possible to replace the background while maintaining shape, composition and visual identity. In subsequent changes, the system should better remember previous decisions and reduce degradation between turns.
OpenAI reports richer textures and greater subject fidelity in uploaded photographs. These are advances declared by the manufacturer; quality and consistency depend on the source material, instructions and number of revisions. The model can still alter unwanted details, misspell text, or produce physically inconsistent elements.
How to use Sketch and image comments
The Sketch function allows you to draw a rough composition directly in ChatGPT and use it as a guide. It is called with @Sketch and can be useful for indicating the arrangement of furniture, the profile of a garment, the position of objects or the framing. A complete textual description then information on style, materials, light and constraints.
Comments can be placed on the image to request a precise change in an area. It’s a flow closer to traditional creative tools than prompting alone. Templates offer starting points for posters, merchandising and product photography, while sharing can include prompting other people to reuse it.
| Function | Novelty |
|---|---|
| Generation | Up to 50% less declared latency |
| Editing | More localized changes |
| Consistency | Improved multi-turn continuity |
| Sketch | Drawing used as reference |
| APIs | Flares and Sunbursts |
| Availability | ChatGPT, Work and Codex |
GPT-Image-2.5 Flare and Sunburst in the API
Two models arrive for developers. GPT-Image-2.5 Flare is the default choice for most applications and prioritizes speed, volume and cost. OpenAI describes it as faster than GPT-Image-2 and suitable for social content, prototypes, visual search and creative products. GPT-Image-2.5 Sunburst instead aims for more rigorous control and premium results, accepting longer generation times.
The distinction helps prevent each request from using the most expensive model. An app can use Flare for drafts and variations, then Sunburst for final assets. Before migrating you need to check prices, formats, limits, data retention and compatibility with your flow. Speed ​​percentage promises are no substitute for a benchmark with real prompts and images.
An effective prompt for Images 2.5
It is best to separate subject, scene, style, composition, lighting and constraints. When editing you need to explicitly indicate what to change and what to keep unchanged. With a reference photo it is useful to define which identifying characteristics should remain, avoiding conflicting instructions. After the first result, a targeted fix works better than a list of disconnected changes.
For a consistent series, lock aspect ratio, palette, framing, and materials before creating variations. Our comparison betweenGPT-6 Astra and GPT-5.6 Solconcerns linguistic models; Images 2.5 instead resolves the visual flow. In complex products the two systems can work together, but costs and checks remain separate.
Limits, provenance and security
OpenAI maintains controls over prompts and images, C2PA metadata, and invisible watermarks. These tools help indicate provenance, but metadata can be lost during recoding or publishing. A plausible image should not be used as evidence of a real event, and portraits require consent and attention to rights.
ChatGPT Images 2.5 seems above all to be an improvement in iterative work: less waiting, more limited changes and greater continuity. The best test is to repeat the same cases used with Images 2.0, measuring how many revisions are needed and how many unwanted parts are changed. It is there, more than in the selected demos, that the real advantage emerges.
We will monitor ChatGPT Images 2.5 over the coming weeks and update this page when independent data, final terms or availability changes emerge. For anyone evaluating it today, the essential rule is to separate accessible features, vendor claims and results that can be reproduced independently.
For further context, read this related CryptoRoad analysis.
