GPT Image 2.5 questions, answered

Twenty questions people ask about this release, answered from OpenAI's own announcement, model pages and system card where possible.

API reference Pricing

Answers below draw on OpenAI's announcement of September 8, 2026, the two developer model pages for Flare and Sunburst, the deployment safety hub system card, and contemporaneous reporting. Where something is not published, this page says so rather than guessing — which happens more often than you would expect for a model this new.

When was GPT Image 2.5 released?
September 8, 2026. OpenAI published the announcement, two API model pages and a system card the same day, and the ChatGPT rollout began immediately.
Is it called GPT Image 2.5 or ChatGPT Images 2.5?
Both, for different surfaces. ChatGPT Images 2.5 is the consumer feature inside ChatGPT; GPT-Image-2.5 Flare and Sunburst are the API model names.
Who can use it?
All ChatGPT tiers, plus ChatGPT Work and Codex, across desktop, mobile and web. Developers reach it through the OpenAI API on the same key as other models.
What replaced what?
GPT Image 2.5 succeeds GPT Image 2, released April 21, 2026. GPT Image 2 itself replaced GPT Image 1.5 and the GPT Image 1 line, with DALL-E 2 and 3 removed from the API in May 2026.
What are the model IDs?
gpt-image-2.5-flare and gpt-image-2.5-sunburst.
What is the maximum output resolution?
OpenAI has not published a new ceiling with this release. The GPT Image line's stated ceiling has been 2K, and the previous generation's API sizes topped out at 1024x1536 and 1536x1024.
Does it support transparent backgrounds?
Yes, and the release specifically improves it. Early adopter Manus called out better transparent-background generation as one of the concrete wins.
How many quality tiers are there?
Six: low, medium, high, xhigh, max and auto. The xhigh and max tiers are new relative to the previous generation's three named tiers.
What is the latency improvement?
Up to 50% lower than Images 2.0 per OpenAI. Manus measured Flare at two to four times the speed of GPT Image 2 on its own evaluation set.
Does GPT Image 2.5 watermark its output?
ChatGPT-generated images have carried C2PA content credentials in this model family. Check the metadata on your own output rather than assuming, since handling differs between surfaces.
Can I use generated images commercially?
Rights follow your OpenAI plan's usage terms, not the model. Review the current terms before shipping generated assets into paid client work.
Is there an official free GPT Image 2.5 generator?
Images 2.5 rolled out to all ChatGPT tiers including free, so a free ChatGPT account is the official free route. Third-party sites offering free access are resellers, not OpenAI.
Does it do video?
No. Both models take text and image input and produce image output. Audio and video are not supported modalities.
What is the Sketch feature?
An in-ChatGPT drawing tool, invoked by typing @Sketch, that lets you draw a reference for the image you want instead of describing it in words.
Are templates available in the API?
No. Templates, Sketch and shareable prompts are ChatGPT interface features. The API exposes the model, not the scaffolding around it.
Which companies shipped with it on day one?
Adobe made the models available in Firefly, and Manus and Higgsfield AI both published evaluation notes alongside OpenAI's announcement.
Is GPT Image 2 still available?
Yes, with no announced deprecation date. GPT Image 1 is the model scheduled to sunset, on October 23, 2026.
Does GPT Image 2.5 pose new safety risks?
OpenAI's system card states neither Sunburst nor Flare crosses the Bio High or Cyber High thresholds in its Preparedness Framework evaluations, and that the safety stack builds on the Images 2.0 foundation.
How do I estimate my token usage?
By measuring. OpenAI notes the GPT Image 2 token calculator does not estimate GPT Image 2.5 consumption, so run representative calls at each size and quality you plan to use.
Where is the official documentation?
OpenAI's announcement post, the two developer model pages for Flare and Sunburst, the image generation guide and the deployment safety hub system card. This site summarises them; it does not replace them.

How this release actually arrived

Worth recording, because the search results from the fortnight before launch are still full of pages insisting the model did not exist. They were correct when written.

  1. August 2026

    Two anonymous checkpoints, mona-lisa-1 and luna-lisa-alpha, appear in public arena blind testing. The community label 'GPT Image 2.5' attaches to them before any official name exists.

  2. September 1, 2026

    A hidden 'Image creation got a major upgrade' modal is spotted in ChatGPT and Codex builds, with its hero asset freshly uploaded.

  3. September 3, 2026

    Independent observers call a release imminent. OpenAI has still announced nothing, and a gpt-image-2.5 model page returns 404.

  4. September 8, 2026

    OpenAI publishes the announcement, two developer model pages, a system card and the ChatGPT rollout, all on the same day. Adobe Firefly ships access simultaneously.

The community guessed the version number correctly and the codenames not at all. Nobody predicted a two-model API split, which turned out to be the structurally interesting part of the release.

Three questions that keep getting answered wrongly

“It is not released yet.” A large number of pages published in the fortnight before launch state flatly that no such model exists, and several of them still rank. They were accurate on the day they were written and are now simply stale. If a page cites a 404 on the model documentation as evidence, check its date before believing it.

“It costs more than the last one.” The token rates did not move. Both new API models bill at the same rates as the model they succeed, which is unusual for a capability upgrade and worth checking against the developer model pages yourself rather than taking anyone's word for it, this site included.

“It does 4K now.” No new resolution ceiling was published with this release. The line's stated maximum has been 2K, and the previous generation's API sizes topped out at 1024×1536 and 1536×1024. If you need native 4K output today, the comparison worth making is against Google's Nano Banana Pro, not against the previous OpenAI model.

What is still genuinely unknown

Per-image costs for the two new quality tiers are not published, and because billing is token-based they cannot be derived from the rate card without measurement. Anyone quoting a firm number for xhigh or max is extrapolating.

Mask behaviour on the edit endpoint has shifted between releases in this family, so inpainting and outpainting workflows carried over from an earlier model deserve a direct test rather than an assumption. The same goes for how content credentials propagate: the chat surface and API responses have not always handled metadata identically.

Finally, independent benchmark placement will take a few weeks to settle. Arena rankings for a model released days ago reflect early, unrepresentative traffic, and the editing arena in particular was already a tight race before this release entered it. Treat any leaderboard position published this month as provisional.

On sourcing: This site is independent and unaffiliated with OpenAI. Where an answer above reflects a vendor claim rather than an independent measurement, it says so. Verify anything load-bearing against OpenAI's own documentation before you build on it.

Keep reading

Features

The nine changes in full, split by model and interface.

How to use it

A ten-step walkthrough inside ChatGPT.

vs Nano Banana Pro

Where Google still wins, by asset type.