Published 9 September 2026
OpenAI released ChatGPT Images 2.5 on 8 September 2026 and says it cuts the wait for an image by up to 50% compared with Images 2.0. The release adds two API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, both charged at $30 per million output tokens, the same rate as GPT Image 2. It is live now for every ChatGPT, ChatGPT Work and Codex user on desktop, mobile and web.
What is new in ChatGPT Images 2.5?
Three things changed in the model itself. It holds on to the subjects in your reference photos better, so a face or a product stays recognisable. It edits only the part you pointed at and leaves the rest alone. And it keeps doing that across several rounds of edits without the picture getting worse each time.
OpenAI also built new features around the model. Sketch lets you draw inside ChatGPT and use that drawing as a reference. Templates cover common formats such as posters and merchandise. You can place comments directly on an image to show what you want changed, and when you share an image you can now share the prompt with it.
Developers get two models instead of one. GPT-Image-2.5 Flare is the fast default. GPT-Image-2.5 Sunburst takes longer and is aimed at precise editing. Both add two quality settings above the old ceiling, xhigh and max, and both accept custom sizes up to 3,840 pixels on the longest edge.
GPT-Image-2.5 Flare and Sunburst pricing and limits
The two new models cost exactly what GPT Image 2 costs, so the case for moving rests entirely on speed and control. Prices are per million tokens, the small chunks of text and image data a model is billed on.
| What is measured | GPT-Image-2.5 Flare and Sunburst | GPT Image 2 |
|---|---|---|
| Image output, per million tokens | $30.00 | $30.00 |
| Image input, per million tokens | $8.00 | $8.00 |
| Text input, per million tokens | $5.00 | $5.00 |
| Highest quality setting | max | high |
| Batch API | Not supported | Supported, $15.00 image output |
| Unsafe images shown in OpenAI’s safety test | 1.09% Sunburst, 1.41% Flare | 1.64% |
These are all OpenAI’s own figures, from its pricing pages and from the safety document it publishes alongside each model. The safety numbers come from a set of deliberately provocative prompts whose size OpenAI leaves blank in the published table. OpenAI marks statistically significant results with an asterisk, and none of the unsafe image figures carries one, so the apparent improvement over Images 2.0 is not solid. API rate limits run from 100,000 tokens and five images per minute on the lowest paid tier up to 8,000,000 tokens and 250 images on the highest.
What OpenAI is not saying about ChatGPT Images 2.5
There is no benchmark table. The launch post carries no image quality evaluation, no win rate and no comparison to any rival. Every claim is measured against OpenAI’s own GPT Image 2. Google, Black Forest Labs and Midjourney are not mentioned once.
Batch processing is not supported on either new model, and GPT Image 2 does support it at half the output price. For anyone generating images in bulk overnight, moving to 2.5 is a straight cost increase. GPT Image 2 also has no announced retirement date, so it is unclear whether the older model is a safe fallback or a dead end.
The safety document is more candid than the launch post. It warns that the model’s higher realism could, without safeguards, make more convincing fake images of real people and events. The launch post does not mention that. It also does not say that the invisible watermark is Google DeepMind’s SynthID, which only the safety document names.
What this means
If you already build on GPT Image 2, move a test workload to Flare this week. The price is identical, so the only question is whether the speed claim holds on your prompts. OpenAI’s own ceiling is 50% lower latency, which is twice the speed. Its customer Manus claims two to four times, so the top of that range is a customer’s number and not OpenAI’s. Teams doing precise brand or product editing, where a logo or a colour has to survive five rounds of changes, should test Sunburst instead and accept the longer wait.
The group that should hold off is anyone pushing images through the Batch API. Losing batch on both new models is a real cost increase with no published timeline for getting it back. For everyone else, the honest summary is that this is a competent point update shipped without evidence. OpenAI released a model it says is better and faster, and published no capability numbers at all to support it. That is a deliberate choice, and it leaves your own test set as the only benchmark that counts.
For more information, visit the official announcement of ChatGPT Images 2.5 on the OpenAI blog.
