It’s 11:40pm, you’re three edits deep into a product mockup, and the model has quietly decided your coffee mug now has two handles. You type “keep everything the same, just change the background.” It changes everything except the background. You close the tab and open Figma like a person who has learned nothing and everything at the same time.
That specific flavor of frustration is what OpenAI says it went after with ChatGPT Images 2.5, released September 8, 2026. The company is calling it its new state-of-the-art image model, and it shipped two companion models to the API alongside it. The headline claim: image generation latency cut by up to 50% compared with Images 2.0.
What actually shipped
Three things worth your attention, and one of them isn’t the speed number.
- Faster generation. Up to 50% lower latency versus Images 2.0. “Up to” is doing work in that sentence, as it always does.
- Sketch. Invoke it with
@Sketchand you can draw directly inside ChatGPT, then use that drawing as a reference for generation. - Two new API models. GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare matches the quality, editing, and speed gains and is the default for most apps. Sunburst adds more precision for detailed editing.
OpenAI also mentions sharper details, more precise editing, and templates for popular formats. And it points at scale, saying people create more than 3 billion images — the framing there is that image generation stopped being a novelty demo a while ago and turned into infrastructure.
Sketch is the interesting one
Speed improvements are table stakes now. Every model release claims them, every release delivers some version of them, and users adapt to the new baseline within about a week. Nobody writes a nostalgic essay about how generation used to take eleven seconds.
Sketch is different because it changes the input channel, not the output quality. Text prompting for images has always had a translation problem: you know exactly where you want the thing, and you’re stuck describing spatial relationships in English like you’re giving directions over the phone. “Slightly left. No, the other left. Smaller. Not that small.”
Drawing a rough box and saying “put it here” skips that entire negotiation. This isn’t a new idea — sketch-to-image conditioning has existed in the open-source tooling world for years, usually behind a wall of nodes and installs. Putting it behind an @ mention inside the chat window everybody already has open is the actual move. Distribution beats novelty most of the time.
The Flare and Sunburst split
Two API models with a stated division of labor is a signal about how OpenAI thinks developers actually use this. Flare is the default: same gains, general purpose, the one you reach for. Sunburst is the precision tier for detailed editing work.
If you’re building anything with an image pipeline, that split matters more than the consumer-facing features. It means you get to pick a cost and latency profile per task instead of routing everything through one model and hoping. It also means you now have a decision to make that you didn’t have last week, and if past patterns hold, plenty of teams will default to the fancier option for jobs that don’t need it and then complain about the bill.
What I can’t tell you yet
I haven’t run these models against a controlled test set, so treat everything below as open questions rather than findings:
- Whether the character and object consistency problem — subjects morphing between edits — is genuinely fixed or just less frequent. OpenAI named it as a target. Named targets and hit targets are different things.
- Whether the 50% latency claim holds on complex multi-subject prompts or only on simple ones.
- What Sunburst costs relative to Flare, and whether the precision gap justifies it for typical editing work.
- How much text rendering has improved, which is the thing designers ask about first and marketing pages mention last.
My read
This looks like a maturity release rather than a leap. Faster, more controllable, better plumbing for developers, and one genuinely useful new input method. Nothing here suggests a different category of capability — it suggests a team working through a list of things users kept complaining about.
That’s not a criticism. The unglamorous fixes are usually what makes a tool go from “impressive demo” to “thing I use on deadline.” Editing consistency in particular has been the wall between AI image tools and real production work, because a model that can’t preserve what you didn’t ask it to change isn’t an editor, it’s a slot machine.
If OpenAI has actually moved that needle, Images 2.5 earns its version number. If it’s moved it partially, you’ll find out the same way the rest of us will — around 11:40pm, on the fourth edit, watching a coffee mug grow a second handle.
🕒 Published: