With ChatGPT Images 2.5, announced 8 September 2026 as OpenAI’s “new state-of-the-art image model”, the image API splits into two named models that charge the same per token. Builders now choose between latency and precision, and price drops out of the decision. And the system card published the same day names the model’s own fidelity as the launch’s new risk.
The API asks how fast, and stops asking how much
GPT Image becomes two models. GPT-Image-2.5 Flare is the fast, everyday generation; GPT-Image-2.5 Sunburst adds extra precision for detailed work at the cost of longer generation times. The model IDs are gpt-image-2.5-flare and gpt-image-2.5-sunburst, each behind a dated snapshot ending 2026-09-08.
The published rates are identical on both pages: $5.00 per 1M text input tokens ($1.25 cached), $8.00 per 1M image input tokens ($2.00 cached), $30.00 per 1M image output tokens, and OpenAI states the rates match GPT Image 2. Choosing between Flare and Sunburst is therefore a latency-versus-precision question. Both models take text and image inputs and produce image output, offer the same quality settings (low, medium, high, xhigh, max, auto) and work either directly in the Image API (generations and edits) or as the image-generation tool’s model in the Responses API.
One footnote cuts the other way: OpenAI says its GPT Image 2 cost calculator does not estimate GPT Image 2.5 token consumption. Cost per image in practice may differ from GPT Image 2 even at matching rates.
What ChatGPT gets is a workflow
Images 2.5 is available to all ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web, and the developer-community announcement says it began rolling out the day of release. The four new creation features change how an image gets made, not how it looks: Sketch turns a drawing made inside the chat into a reference; templates cover popular formats such as flyers and product photos; comments anchor to a part of an image so an edit targets that region; and prompt sharing lets someone else rerun an idea with their own inputs.
The release notes carry the fine print: on mobile, Sketch is invoked by typing @ in the message box and selecting it, and full-screen images can be edited or annotated with comments; in The Verge’s hands-on, @Sketch opened a drawing window on desktop as well. Templates are not in Work mode yet, and existing image-generation limits are unchanged. While an image renders, some users get animated dots and a playable Snake game.
The performance numbers are OpenAI’s
The company claims latency reduced by up to 50% compared with Images 2.0, and says 2.5 delivers more natural lighting, richer textures, better preservation of a reference photo’s subject and more reliable multi-turn editing. The announcement publishes no measurement method: no resolutions, quality settings, regions or sample sizes. No third-party benchmark for Images 2.5 existed as of 9 September 2026. The framing number is self-reported as well: more than 3 billion images created every week across ChatGPT Images and the API’s GPT-Image models is OpenAI’s figure, with no disclosed method.
A secondary line of reporting positions Flare as “the default choice for most applications”, making higher-quality images than GPT-Image-2 at 50% lower latency. That wording appears in Unite.AI’s account but was not visible in the portion of OpenAI’s announcement the research fetched, and the report drops the announcement’s “up to” qualifier, so it is recorded here as reported rather than confirmed.
The system card treats the improvement as the risk
The most interesting document of the release concedes the selling point and then inverts it. The system card, also dated 8 September, states that the model improves infographic accuracy and layout, and names heightened realism versus Images 2.0 as a new safety challenge: potentially more convincing deepfakes, including political, sexual or otherwise sensitive imagery of real people, places or events.
Against that risk the card describes a stack of defences, layer by layer: LLM-based upstream refusals, safety classifiers on input images, a combined prompt-plus-image analysis against malicious edits, and a policy check on the generated image before it is shown.
Provenance is the part that does not change, and it reads differently next to the risk the card itself names. OpenAI says it will “continue to use C2PA metadata and invisible watermarking to help identify images made with our tools”: the same measure as Images 2.0, with nothing new announced for 2.5. The model raises the realism and the label that lets you recognise its output stays where it was, and that label only works where something reads it: metadata is gone the moment an image passes through a platform that strips it.
What to watch
- The price is neutral; the segmentation is not yet documented. Beyond Sunburst trading time for precision, the model pages give the two identical rates, settings and endpoints. What the extra time actually buys is undocumented.
- The changelog did not notice the release. As fetched on 9 September, the API changelog has no September 8 entry: its September begins with GPT-6 Astra on Sep 3 and never lists the two image models. Pinned integrations will learn about snapshots from the model pages, not the changelog.
- Nobody says the older models are gone. The announcement is silent on whether Images 2.0 or the older DALL·E flows retire in ChatGPT, and the retirement claim circulating in press coverage appears in no OpenAI source.
Further Reading
- Introducing ChatGPT Images 2.5 — OpenAI’s announcement: model, features and the two API models
- ChatGPT Images 2.5 System Card — the deepfake risk framing and the layered safety stack
- Image generation guide — how to select Flare or Sunburst in the Image API and the Responses API