ChatGPT Images 2.5: What Actually Changed for Photos of Yourself

OpenAI released ChatGPT Images 2.5 on 8 September 2026, along with two API models and a drawing tool. Most coverage will lead with the headline quality claims. For anyone generating photos of themselves, one clause in the announcement matters more than the rest of it combined:

"Better preserves subjects from users' reference photos."

That is the entire problem this site exists around. Everything else is a bonus.

What Actually Shipped

Three things, and they are easy to confuse because the naming is close.

ChatGPT Images 2.5 — the consumer model, inside ChatGPT. OpenAI calls it their new state-of-the-art image model. It is rolling out to all ChatGPT, ChatGPT Work and Codex users across every tier, on desktop, mobile and web.

GPT-Image-2.5 Flare — an API model, positioned as the default choice for most applications. Higher quality than GPT Image 2 at 50% lower latency.

GPT-Image-2.5 Sunburst — the premium API option, built for workflows where editing precision matters most.

If you use ChatGPT directly, you have Images 2.5 and the other two names are irrelevant to you.

The Reference-Photo Claim

The stated improvements are more natural lighting, richer textures, better handling of complex subjects, and better preservation of subjects from reference images. It also holds a visual style across an extended conversation rather than drifting.

Why that last one matters as much as the first: building a set of images of yourself has never failed at image one. It fails at image four, when the face has quietly become someone else's. A model that holds style and subject across a long conversation is describing exactly the failure covered in keeping AI photos consistent across generations.

A caveat worth stating plainly. "Better preserves subjects" is OpenAI's own description on launch day. It is not a measured result, and nobody outside OpenAI has tested it against a hard identity benchmark yet. Treat it as a promise worth testing, not a finding. The way to test it is boring and effective: run the same reference photo and the same prompt through it and through whatever you used last week, generate six images each, and look at image six.

Sketch Is the Genuinely New Thing

The feature nobody will lead with, and the one that changes how you prompt: type @Sketch in ChatGPT and draw directly, as a visual reference for the generation.

Composition is the hardest thing to write in words. "Subject at the left third, looking out of frame, horizon low, car entering from the right edge" is a sentence you have to construct carefully and the model still has to interpret. A thirty-second scribble says it exactly.

This does not replace prompt language — it replaces the part of prompt language that was always doing a bad job. Light, lens, surface behaviour and mood still have to be written, and they still transfer between models. Layout no longer has to be.

There are also templates for common formats like flyers and product photos, and you can now comment directly on a generated image to request targeted edits rather than describing the region in prose.

Where This Leaves the Comparison

Our model comparison for photos of yourself recommended GPT Image 2 as the default for single strong images of a person. Images 2.5 supersedes it on OpenAI's own account, and Flare claims higher quality at half the latency in the API.

The comparison's other recommendations are untouched. FLUX.2 still takes ten simultaneous reference images, which is still the answer for a consistent set. Gemini still wins on conversational iteration and SynthID provenance. MAI-Image is still the repair tool — and worth noting, MAI-Image-2.6-Preview has reportedly closed most of the gap at the top of the blind-preference arenas, which was a 21-point margin at last reading.

The honest summary: the top of this field changed twice in five days, and the decision framework did not.

What Does Not Change

This is the fourth model release in six months and the advice has survived all four, because it was never about the model:

- Fix the reference photo before you change tools. Even, diffuse light. Neutral expression. Face square to camera. No sunglasses, no beauty filter. A beautified reference bakes the beautification in, and every model faithfully reproduces a face that is already slightly fictional. This is a larger quality gain than any model switch.
- Write in camera language. Lens, light position, falloff, surface behaviour. It means the same thing to every model in this list and it will mean the same thing to whatever replaces them.
- Identity never transfers. A reference is bound to a generation, not to a prompt. Moving to Images 2.5 means re-establishing your face there.

Every prompt in the archive is written that way deliberately. It is why they still work after a week in which the leader changed twice, and the Virtual Photoshoot runs from your own reference upload for the same reason.

*Released 8 September 2026. Performance claims here are OpenAI's own launch-day descriptions and have not been independently verified.*