On 8 September 2026, OpenAI introduced ChatGPT Images 2.5, a new image model it describes as state of the art, alongside two API models and a set of creative controls inside ChatGPT. The company says the model produces sharper details, more natural lighting and richer textures, and follows editing instructions more reliably across multiple turns.
The headline number is speed. According to OpenAI, image generation latency is reduced by up to 50% compared with Images 2.0. The company also says more than 3 billion images are created every week across ChatGPT Images and the GPT-Image models in the API.
Images 2.5 is rolling out today to ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web. In the API, OpenAI is releasing GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
Confirmed
OpenAI's own description of the model centers on four areas: fidelity when working from reference photos, precision editing, multi-turn consistency, and instruction understanding.
On fidelity, the company says Images 2.5 is better at transforming familiar subjects across new settings, visual styles, and compositions, with subjects looking more recognizable and distinctive features more likely to carry through. For API builders, OpenAI frames this as making reference-led workflows more reliable, so variations stay anchored to the original source.
On editing, the model is described as better at changing only what was asked while keeping the rest of the image the same, even with complex subjects and backgrounds. OpenAI says developers can update a single element — a product, background, or piece of copy — while preserving the subject, composition, and brand treatment around it.
On multi-turn work, OpenAI says earlier changes are more likely to stay consistent during longer conversations and that each new edit builds on prior work without degrading image quality over time. The company positions this as relevant to production workflows where targeted changes are needed without rebuilding an entire asset.
The two API models split by use case. GPT-Image-2.5 Flare, according to OpenAI, brings the same quality, editing, and speed improvements and is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency. GPT-Image-2.5 Sunburst is aimed at premium visual workflows that need tighter control across edits, such as production-ready campaign creative or polished product imagery, and comes with longer generation times.
On the ChatGPT side, OpenAI is adding Sketch, which lets users draw directly in ChatGPT as a reference for the final image, invoked by typing "@Sketch". Templates offer starting points for common formats such as posters, flyers, product photos, and merch. Users can also place comments directly on images for focused editing, and can choose to share the prompt behind an image so others can run the same idea with their own photos and details.
On safety, OpenAI says Images 2.5 builds on existing safeguards with checks on prompts and images, and that it continues to use C2PA metadata and invisible watermarking to help identify images made with its tools. The company points to a system card for its evaluations and approach.
Unknown
Most of the performance claims here come from OpenAI's own reporting, and the material does not include independent replication of the fidelity, editing-consistency, or latency figures. The "up to 50%" framing is a ceiling, not a typical result, and OpenAI does not break the number down by workload, resolution, or model variant in the announcement.
Pricing is also unresolved in the source. OpenAI says GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare are available in the API and links to its business pricing page, but the announcement text does not carry per-image rates. For teams budgeting high-volume generation, that gap matters more than the latency claim.
Sunburst's trade-off is stated but not quantified: longer generation times for extra precision, with no figures on how much longer. And while the model is described as available to all ChatGPT, ChatGPT Work, and Codex users across all tiers, the announcement does not detail whether every tier gets the same generation limits or the same access to Sketch, templates, and image comments.
Our take
The interesting shift is not the latency number but where OpenAI is putting control. Sketch, templates, on-image comments, and prompt sharing are all attempts to move the creative loop inside ChatGPT rather than into a separate editor. That is a distribution play as much as a model release: if the reference image, the edit instruction, and the revision history all live in one conversation, switching to a rival image tool costs the user more than a model swap. Whether that holds depends on the pricing page, which the announcement leaves blank.
