ChatGPT Images 2.5: what one image really costs
OpenAI released ChatGPT Images 2.5 on 8 September 2026 with two API models. The quality labels moved: medium now costs a quarter of what it used to.
OpenAI released ChatGPT Images 2.5 on September 8, 2026. This post sticks to what OpenAI published, with the numbers taken from its own pricing calculator, and then explains the one thing most coverage got wrong.
The names
The consumer product in ChatGPT is called ChatGPT Images 2.5, with Images plural. The API ships two separate models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, addressed as gpt-image-2.5-flare and gpt-image-2.5-sunburst. There is no single product called "GPT Image 2.5" in OpenAI's own copy, even though almost everyone writes it that way.
Flare is the one OpenAI calls the default for most applications: better images than GPT Image 2 at half the latency. Sunburst trades generation time for tighter control across successive edits, aimed at campaign creative and polished product imagery.
What changed
- Better fidelity when working from reference photos, so subjects stay recognizable across new settings and styles.
- Edits that change only what you asked for while the rest of the frame holds.
- Multi-turn consistency, so a long editing conversation stops degrading the image.
- More complex layouts, including transparent backgrounds.
- Generation latency down by up to 50% against Images 2.0.
In ChatGPT itself the release added Sketch, which lets you draw a rough guide and have the model finish it, plus format templates, comment-based edits placed directly on an image, and the option to share the prompt alongside a shared image.
Images 2.5 rolled out to every ChatGPT tier plus ChatGPT Work and Codex on the day of release. OpenAI applies C2PA metadata and invisible watermarking to images made with its tools.
What one image costs
OpenAI bills image generation by token: $30 per million image output tokens, $8 per million image input tokens, $5 per million text input tokens. Per image, at 1024x1024, that works out to:
| Quality | GPT Image 2.5 | GPT Image 2 |
|---|---|---|
| low | $0.00588 | $0.00588 |
| medium | $0.01317 | $0.05268 |
| high | $0.05268 | $0.21072 |
| xhigh | $0.09366 | not available |
| max | $0.21072 | not available |
Those are image output tokens only. Input tokens and streamed partial images bill on top, and a bigger canvas costs more: the same medium setting at 2048x2048 is $0.02676.
The part worth knowing
Look down the two columns and the pattern is not "2.5 is cheaper". The quality labels moved.
Medium on 2.5 costs a quarter of medium on GPT Image 2. High on 2.5 costs exactly what medium used to cost. The new max setting costs exactly what high used to cost. The old ceiling became the new middle, and two settings were added above it.
This matters when you read a product page. A tool that announces it upgraded to 2.5 has told you almost nothing, because the same model at low and at max differs by a factor of thirty-six in what the image costs to produce. The setting is the price, not the model name. When a service will not tell you which quality it runs, that silence is the answer.
It also means resellers can quote wildly different prices for the same claimed model and all be telling the truth. Before you compare two credit prices, check that both are talking about the same quality setting.
Where we stand
Image 2.5 is our product name, not a model name, and we are not affiliated with OpenAI. We name every model in the workbench in full, with its API model id and its credit price per image, precisely because the paragraph above is so easy to hide behind.
Sources: OpenAI, Introducing ChatGPT Images 2.5, and the OpenAI API pricing documentation, both read September 9, 2026.