OpenAI ships ChatGPT Images 2.5 with Sketch and up to 50% lower latency

OpenAI shipped ChatGPT Images 2.5 on 8 September 2026, and it arrived with a drawing pad attached. Type @Sketch into the chat box and a canvas pops open, and then you tell ChatGPT what to make of whatever you scribbled. The Verge tested it with a cat doodle drawn using a computer mouse, and got a realistic photo back.

The model changes sit underneath that party trick. OpenAI says Images 2.5 makes images with “more natural lighting and richer textures”, follows editing instructions better “across multiple turns”, and has cut image generation latency by “up to 50% compared with Images 2.0”, per the same report. OpenAI’s announcement describes it as a way to turn “ideas, sketches, and reference photos” into more polished images.

That last phrase carries a second claim, which is that the model “is better at preserving the subjects in your reference photos”. Simon Willison quoted that line from the post on the day of the release. It’s a statement about how a model reads an uploaded image, so it isn’t something a demo settles either way.

Willison also pulled the number that puts the launch in proportion. OpenAI’s image models are used for “more than 3 billion images across ChatGPT Images and the GPT-Image models in the API”, he quoted from the same post. At that volume a latency cut reads as a serving-cost decision as much as a product one.

More than 3 billion images across ChatGPT Images and the GPT-Image models in the API.

OpenAI, quoted by Simon Willison

Still, none of it is gated to a preview list. Images 2.5 is out now for ChatGPT, ChatGPT Work and Codex users on desktop, mobile and web, and developers get two model IDs rather than one. OpenAI’s image generation guide says to “Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation”.

Model ID	OpenAI’s positioning	Default snapshot	Image output
gpt-image-2.5-sunburst	Most capable for image generation and editing	gpt-image-2.5-sunburst-2026-09-08	$30 per 1M tokens
gpt-image-2.5-flare	Fast, high-quality everyday generation	gpt-image-2.5-flare-2026-09-08	$30 per 1M tokens

The pricing didn’t move, which is the more interesting half of that table. Both models list text input at $5 per million tokens, image input at $8 and image output at $30, and the docs say flatly that “Token rates match GPT Image 2”. So the quality and latency changes land without a token price rise, which is worth noting in a year of three-week release cycles.

Most of the new surface area, though, is in what happens after the image appears. ZDNet’s tester found built-in templates for bobblehead, statue, anime, sticker and icon, plus markup, comments, background removal and resizing before you download a PNG. Their drawing of a person in a hat came back as a turkey wearing a Pilgrim hat, which they took as a comment on their drawing.

The ceiling sits in the rate limit table rather than the benchmark copy. OpenAI’s docs put Tier 1 API accounts at 5 images per minute and Tier 5 at 250, and list image generation and image edit as the supported endpoints. Watch that images-per-minute column before planning a batch pipeline, because at five a minute the constraint is how often you’re allowed to ask, not what the model can draw.