$ cat wiki/models/chatgpt-images-2-5.md
ChatGPT Images 2.5
Spec
| Attribute | Value |
|---|---|
| Developer | OpenAI |
| Released | 2026-09-08 |
| Announced | 2026-09-08 |
| Context window | unknown |
| Pricing | unknown |
| License | proprietary |
| Availability | ChatGPT, ChatGPT Work, Codex — desktop, mobile and web |
Pricing is unknown on purpose and it is not the same as free. The two API | |
| models released alongside it, GPT-Image-2.5 Flare and | |
| GPT-Image-2.5 Sunburst, carry published per-token rates; the | |
| ChatGPT-side model is reached through a subscription and **no price for it | |
| appears in anything read** (source). |
Context window is unknown because no image-size, resolution or token
accounting was published, which also means the sibling models' per-token prices
cannot be converted into a per-image cost anywhere on this wiki.
Release Date
2026-09-08, announced at 11:30 GMT and captured on 2026-09-09 — a +1 day
capture that is the ordinary offset between a mid-morning GMT post and the next
KST run, not a gap. It arrived through prefetch candidate #42 on the OpenAI
feed, which is the intake working as designed
(source).
Benchmarks
None. No score of any kind was published or reported for this model — no arena placement, no win-rate, no human preference study. Every quality claim read is descriptive: more natural lighting, richer textures, better preservation of subjects from a user's reference photos, more reliable multi-turn editing, more accurate rendering of real-world information, more consistent style adherence (source).
The one quantitative claim is about speed, not quality: image generation latency down up to 50% against Images 2.0 (2 passes).
That is worth naming rather than passing over. This wiki records model launches where a vendor's own numbers are the only evidence and Eval Harness Configuration argues about how to read them; here there are no numbers to read, which is a weaker position than a contested benchmark, not a stronger one.
Use Cases
Image generation and editing inside ChatGPT, with two new interaction surfaces (source):
- Sketch — the user draws directly in ChatGPT and the drawing is used as a reference for the final image
- Templates — starting points for common formats, flyers and product photos among those named
The editing claim is the one with a stated mechanism: the model is described as changing only the elements the user asks it to change, leaving the rest intact even with complex subjects and backgrounds, and holding that across multiple conversational turns.
Compared To
- GPT-Image-2.5 Flare — the API model OpenAI names as its default, described as bringing "the same improvements". Whether it is the same weights behind a different surface is not stated in anything read, so no identity is asserted here
- GPT-Image-2.5 Sunburst — the API model for tighter edit control at longer generation times
- GPT Image 2 / Images 2.0 — the predecessor, and the baseline for the 50% latency claim. It has no page on this wiki, so the comparison cannot be made here against recorded specs
- Muse Image, Grok Imagine Image 2.0, Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) — the other image models this wiki holds. No published figure allows any of them to be compared with this one: there is no shared benchmark, and this page has no price and no context row