OpenAI released ChatGPT Images 2.5 on September 8, 2026, and paired the new image model with a drawing tool called Sketch that lets users doodle inside ChatGPT and use the result as part of a prompt. OpenAI says the updated model cuts image generation latency by up to 50% compared with Images 2.0, produces "more natural lighting and richer textures," and is better at following editing instructions "across multiple turns." The rollout covers ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.
Sketch activates by typing @Sketch into the chat box, which opens a drawing window inside the interface. The user draws whatever they want, then writes a prompt telling ChatGPT how to render it. A rough mouse-drawn scribble of a cat, per the tool's first public test, came back as a realistic photo matching the accompanying instructions.
The second new interaction pattern is inline commenting on generated images. Instead of writing a new prompt describing everything that should change, users can click on a specific region of an image and leave a note — "make the eyes green instead of blue," for example — and the model edits just that area on the next turn.
Key facts
- 01OpenAI released ChatGPT Images 2.5 on September 8, 2026, alongside a new Sketch tool for doodle-based prompts.
- 02Image generation latency is down up to 50% compared with the prior Images 2.0 model.
- 03Sketch activates by typing @Sketch in the chat box, opening a drawing window inside ChatGPT.
- 04Users can now leave comments directly on parts of an image to request targeted edits.
- 05The release covers ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web.
The 50% latency claim, if it holds up in practice, is the more consequential number here. Image generation in ChatGPT has historically been the slowest part of the assistant's multimodal loop, often taking tens of seconds per image and breaking the flow of iterative editing sessions. Cutting that time roughly in half brings the interaction closer to a real design conversation rather than a batch job.
The "multiple turns" language is also load-bearing. Earlier ChatGPT image models struggled to preserve subject identity and composition across edits — ask for a small change and the whole scene would drift. Images 2.5's pitch is that a chain of small refinements now stays coherent, which is what makes region-level commenting useful in the first place.
Sketch itself is a smaller but strategically pointed feature. Doodle-to-image workflows have been a specialty of standalone tools and design-focused startups; folding one directly into ChatGPT reduces the reason to leave the assistant for a dedicated app. It also lowers the prompt-engineering skill floor — a bad drawing communicates spatial intent that a paragraph of text often cannot.
Availability spans the paid and developer surfaces at once. ChatGPT Work users get it as part of their subscription, and Codex users — OpenAI's coding-focused environment — get access on the same day, which suggests OpenAI wants the model treated as general infrastructure rather than a consumer novelty.
OpenAI did not publish benchmark comparisons against competing image models from Google, Black Forest Labs, or Midjourney, so the Images 2.5 improvements are described only relative to the company's own prior version. The latency figure is also framed as an upper bound — "up to 50%" — meaning real-world speedups will vary by prompt complexity and load.
Rapid model iteration on the image side is a shift in emphasis for OpenAI, which spent most of the past year pushing reasoning and agent capabilities in text. Images 2.5 arriving with a UI feature attached — Sketch and inline comments — rather than as a silent model swap suggests the company sees creative tooling as a product surface worth defending, not just a checkbox next to text generation. The competitive read is that the assistant war is broadening back out to multimodal fluency, and whoever makes image editing feel like a conversation rather than a lottery wins the next round of daily use.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
