What if the biggest problem with your image generation feature was never the image quality?
I build bots for a living. Discord bots, Slack bots, little internal tools that turn a text prompt into a product mockup. And in almost every one of those projects, the complaint I get from users is not “this picture looks wrong.” It’s “why is this taking so long?” A user types a prompt, watches a spinner, and by second twelve they’ve switched tabs. By second twenty they’ve forgotten they asked.
So when OpenAI released ChatGPT Images 2.5 on September 8, 2026, the number I cared about was not the sharper details. It was the latency cut of up to 50% compared with Images 2.0. That’s the difference between a bot that feels like a conversation and a bot that feels like a fax machine.
What actually shipped
OpenAI is calling Images 2.5 its new state-of-the-art image model, with sharper details and more precise editing. Alongside it, two companion models landed in the API:
- GPT-Image-2.5 Flare — matches the quality, editing, and speed gains, and is the default for most apps.
- GPT-Image-2.5 Sunburst — adds more precision for detailed editing work.
On the ChatGPT side, there’s a new Sketch feature invoked with @Sketch that lets you draw directly in the chat. Your drawing becomes a reference for generation. OpenAI also added templates for popular formats.
The reporting frames Images 2.5 as targeting the three most common complaints about Images 2.0: slow generation, subjects morphing between edits, and imprecise editing. That framing tells you something. This is a maintenance release wearing a new version number, and I mean that as a compliment. Maintenance releases are what production systems actually need.
Why two API models is the interesting part
A single model forces you into a single tradeoff. Two models let you route.
Here’s how I’m thinking about the architecture for anything I ship this quarter. Flare is the default path. It handles the high-volume, user-facing requests where speed matters and the output is going to be glanced at, reacted to, and possibly regenerated. Sunburst is the escalation path. When a user says “no, move the logo left and keep everything else identical,” that’s a precision editing job, and precision is what Sunburst is for.
That’s a routing layer, not a model choice. Something like:
- First generation from a text prompt → Flare
- Casual iteration, vibe checks, thumbnails → Flare
- Targeted edits on an approved image → Sunburst
- Anything where the user explicitly asked for detail work → Sunburst
You build this as a small decision function in front of your image service, and you log which path each request took. Six weeks later you’ll know whether your users are mostly generating or mostly refining, and that tells you more about your product than any survey will.
The morphing problem was the real bug
Of the three complaints Images 2.0 reportedly had, subject drift between edits was the one that killed features for me. You generate a character. The user likes the character. They ask for one change. You get back a different character wearing the requested change.
That breaks any workflow built on iteration, which is most useful workflows. If Images 2.5 holds subjects stable across edits, a whole category of bot features becomes buildable that previously wasn’t: character sheets, product variant generators, storyboard tools where panel three has to look like panel one.
I’d treat this as the thing to test first. Generate a subject, run five sequential edits, and compare frame one to frame five. That single test will tell you more about whether 2.5 changes your roadmap than any benchmark chart.
Sketch is a UX signal, not just a feature
Sketch is a ChatGPT feature, invoked with @Sketch, not something you get in the API for free. But it’s worth watching as a design pattern.
Drawing a rough shape is a faster way to express spatial intent than writing a paragraph about spatial intent. “Put the cat on the left, slightly behind the lamp, facing away” is four seconds of scribbling and thirty seconds of typing. If OpenAI is investing in sketch-as-reference, the signal is that image models are getting better at reading structural hints, which means multimodal reference inputs are worth designing for in your own tools.
What I’d do this week
Swap Flare in as your default and measure your actual latency change against your own workloads, not the headline number. Add a routing flag for Sunburst on edit operations. Run the five-edit drift test. Then decide whether the speed gain lets you turn on features you’d previously shelved as too slow to be usable.
Faster and steadier is not a flashy pitch. It’s just the version of the feature people will actually finish using.
🕒 Published: