OpenAI's ChatGPT Images 2.5 Adds Sketch-to-Image Tool

What OpenAI announced

OpenAI rolled out ChatGPT Images 2.5 on September 8, introducing a new image model alongside a set of interface features built around a single idea: let people show what they want instead of just describing it 17. The headline addition is Sketch, a drawing canvas built directly into ChatGPT. Type “@Sketch,” doodle a rough layout — a room, an outfit, a cartoon cat — add a few words about style or materials, and ChatGPT turns the drawing into a finished image 14812.

OpenAI's own announcement frames Images 2.5 as a broad quality upgrade: more natural lighting and richer textures, better preservation of subjects pulled from reference photos, more reliable editing across multi-turn conversations, and generation latency cut by up to 50% compared with Images 2.0 7. That latency figure is repeated almost verbatim across coverage — Yahoo Tech, TechRepublic, Android Headlines, PCMag and The Verge all cite the same 50% speed claim, sourced to OpenAI 245810. Beyond Sketch, the release adds Templates for common formats like posters and merchandise, the ability to drop comments directly on parts of an image to request localized edits, and a prompt-sharing feature so someone else can reuse your exact creative recipe with their own photos 457. Developers get two new API models: GPT-Image-2.5 Flare, a faster default tuned for high-volume and everyday generation, and GPT-Image-2.5 Sunburst, a slower model aimed at precision editing work 711. Simon Willison's technical read of the release concludes Sunburst is the stronger option specifically for editing-heavy workflows, echoing OpenAI's own positioning of the two models 11.

OpenAI says the scale behind this update is substantial: more than 3 billion images are generated weekly across ChatGPT Images and the GPT-Image API 7. The rollout covers ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web, on all tiers 4578.

Hands-on reactions

The Verge tested Sketch directly, turning a mouse-drawn cat doodle into a realistic image and then using the new comment tool to change the cat's eye color — a small but telling demonstration of how the localized-editing feature is meant to work in practice 8. TechRadar spent 24 hours with the update and singled out Sketch as one of three features people will actually use, describing the drawing tool as functionally similar to a stripped-down MS Paint, good enough for a quick sketch by finger or mouse before handing it off to the model 12. TechRepublic framed the overall shift as moving from a “describe until it's right” loop toward a “point and adjust” workflow 4, a framing that lines up with how OpenAI itself describes precision editing: changing one element — a product, a background, a line of text — without disturbing everything around it 7.

Context: this follows Images 2.0

Images 2.5 builds directly on ChatGPT Images 2.0, launched in April 2026, which introduced web-search-informed "thinking" capabilities, up to 2K resolution, aspect ratios from 3:1 to 1:3, generation of up to eight consistent images from one prompt, and improved multilingual text rendering 1617. NewsBytes explicitly frames Images 2.5 as building on that April predecessor 6. The naming convention — a "2.5" rather than a full version jump — undersells how much the product layer has changed: Sketch, Templates, comments and prompt-sharing turn ChatGPT into something closer to a lightweight creative workspace than a single prompt box.

Safety and provenance

OpenAI's system card for Images 2.5 acknowledges the obvious tension in a more capable, more photorealistic editing tool: heightened realism could, without safeguards, enable more convincing deepfakes of real people, places or political events 9. The company describes a layered safety stack — upstream prompt checks, a multimodal safety reasoning model that screens both inputs and outputs, and continued use of C2PA metadata plus a newly added invisible watermarking layer via Google DeepMind's SynthID 9. In OpenAI's own adversarial testing, designed specifically to provoke policy-violating outputs, the rate of unsafe images that slipped through was 1.09% for Sunburst and 1.41% for Flare, compared with 1.64% for Images 2.0 — though OpenAI cautions this test set does not reflect normal usage patterns, and that none of the differences reached its stated statistical-significance threshold 9.

Where the reporting agrees

Across Digital Trends, Yahoo Tech, TechRepublic, Android Headlines, NewsBytes, The Verge, TechRadar and PCMag, the factual core of this release is remarkably uncontested 1245681012. Every outlet that covers the launch confirms the same basics: a new model called ChatGPT Images 2.5, a Sketch feature invoked with "@Sketch," roughly 50% faster generation than Images 2.0, and availability across ChatGPT, ChatGPT Work and Codex on desktop, mobile and web. Multiple outlets independently tested or described the comment-based editing feature the same way — pointing to a specific region of an image to request a change — which suggests this is a genuinely distinct addition rather than a repackaged version of prior editing tools.

Where it doesn't

The disagreements here are less about facts and more about framing and scope. PCMag places the release in a competitive timeline, noting it arrived "just days after OpenAI released GPT-6 Astra" and alongside a separate feature connecting third-party messaging services to ChatGPT — context none of the other outlets mention 10. That timeline detail is worth treating as PCMag's own reporting rather than a corroborated industry fact.

A more significant divergence concerns scope creep beyond the actual announcement. One source discusses a broader wave of anxiety about AI models acting autonomously and breaching systems, a topic entirely unrelated to image generation 3. That piece does not describe Images 2.5 at all and should not be read as commentary on this release; it reflects a separate strand of AI-safety discourse that happened to surface around the same period.

Similarly, reporting on Sora — OpenAI's video-and-audio generation model — and a since-unconfirmed report that Sora tools might be folded into ChatGPT's main interface 1315 describe a different product line. OpenAI's own Images 2.5 launch page includes demo "videos," but explicitly states these are composites of multiple still images illustrating editing consistency, not evidence that Images 2.5 generates video 7. Treating Images 2.5 as a step toward native video inside ChatGPT would overstate what OpenAI has actually confirmed; the PCMag report on Sora-in-ChatGPT was explicitly sourced to unnamed insiders via The Information and remains unconfirmed by OpenAI 15.

Finally, a survey of competing video models — Veo 3.1, Seedance 2.0, Runway Gen-4.5, Kling 3.0 and others — shows the wider industry moving toward reference-image and multi-input control rather than longer text prompts 14. That trend runs parallel to what Sketch does for still images, but it is an industry-wide pattern documented separately, not something Images 2.5 itself claims to address.

The most defensible reading

The evidence supports treating Images 2.5 as exactly what OpenAI calls it: an incremental model upgrade wrapped in a more consequential interface change. The consistent 50% latency figure, repeated by outlet after outlet without independent verification, is a company claim measured against OpenAI's own prior model — none of the coverage cites third-party benchmarks against Midjourney, Google, or Black Forest Labs, so it should be read as a vendor-reported number rather than an industry-verified one.

What's better supported is the qualitative shift the hands-on reviews converge on: Sketch and inline comments meaningfully reduce the friction of getting an AI image to match what's in someone's head, by letting geometry be drawn rather than described. That's a real product change, confirmed by independent testing at The Verge and TechRadar, not just a press-release claim. The connection to AI video generation, by contrast, is inference rather than fact — Images 2.5 is a still-image release, and any link to Sora or native video inside ChatGPT remains speculative pending OpenAI's confirmation.