GPT Image release timeline
Verified against OpenAI documentation on 2026-09-09
GPT Image 2.5 is the latest entry in a line that runs back through GPT Image 2 and GPT-Image-1.5. This timeline lists what shipped on each date, and it separates confirmed releases from the community reports that circulated before the last one. Verified against the OpenAI documentation on 2026-09-09.
How to read this timeline
The entries below run oldest first. Each one is either a dated OpenAI release or a community report from before a release, and every community item says so in its own text. That separation matters here, because the weeks before ChatGPT Images 2.5 produced a lot of confident writing about models nobody had confirmed.
Only two kinds of things appear. Confirmed releases carry a date OpenAI published and describe what actually shipped. Community items carry the date the report surfaced, not a release date, and none of them were ever acknowledged. Nothing on this page predicts a future release, and nothing here should be read as a roadmap.
Release history
GPT-Image-1.5
GPT-Image-1.5 arrived as the first release in the naming scheme that the later models follow. It set the pattern the family has kept since: one image model family serving both the ChatGPT product and the developer API, with the same generation and editing endpoints behind both. Anyone still calling this model today is on the oldest of the three generations covered here. The migration path forward is mostly a change of model string, since the endpoints and the request shape did not change underneath it. Prompts do need retesting after a move, because the newer models weigh context differently and wording tuned here does not always transfer cleanly.
GPT Image 2 and ChatGPT Images 2.0
GPT Image 2 shipped alongside ChatGPT Images 2.0, pairing an API model with the product experience built on it. It supports up to 16 reference images in a single edit request, which made it the first of these models practical for work where a subject or a product had to stay consistent across a set of pictures. Its token rates are the ones GPT Image 2.5 inherited unchanged, so the newer generation did not raise the price of anything. Images 2.0 also became the baseline that later speed and safety comparisons measure against, which is why it keeps appearing in the numbers published for Flare and Sunburst.
Two anonymous arena checkpoints, community speculation
Community reports, which are speculation and not confirmed fact, described two anonymous evaluation-arena checkpoints named mona-lisa-1 and luna-lisa-alpha appearing during August 2026. People testing them concluded they were an unreleased image model and connected them to OpenAI. OpenAI has never said what those checkpoints were, and no release note mentions either name. They are listed here because they shaped what people expected from the next release, not because they form part of the record. Treat any article that presents them as an official preview as unverified, including the performance claims that were attached to them at the time.
A hidden upgrade modal in ChatGPT, community report
A community report, again speculation rather than confirmed fact, described a hidden modal found in ChatGPT behind a disabled feature flag. The text people reported reading was that image creation got a major upgrade. Finding an unreleased string behind a flag is not an announcement, and OpenAI made no statement about it. What the report did establish for the people watching was timing, since a disabled flag with finished copy usually sits close to a launch. The release followed within the week, which is why this item is worth keeping in the record even though it was never official.
ChatGPT Images 2.5 with Flare and Sunburst
OpenAI released ChatGPT Images 2.5 to ChatGPT, ChatGPT Work and Codex users at every tier, on desktop, mobile and web. The API gained two models, `gpt-image-2.5-flare` as the fast default and `gpt-image-2.5-sunburst` for editing precision, each with a dated snapshot carrying the release date. The improvements OpenAI named are more natural lighting, richer textures, better preservation of reference-photo subjects and more reliable instruction following across multiple editing turns. ChatGPT also gained Sketch, Templates, comments pinned onto an image, sharing an image with its prompt, and version history per edit. Token rates matched GPT Image 2, so nothing about the release changed what an image costs to make.
What the cadence suggests if you build on these models
Three generations have shipped inside a year, and each one arrived without a public preview period. There is no way to read a next date out of that, and this page will not try. What the pattern does tell you is that a model you are calling today may not behave identically in a few months, and that the gap between announcement and general availability has been effectively zero.
The practical response is to pin a dated snapshot anywhere output consistency matters. A snapshot holds the version you tested against. The floating alias moves when OpenAI updates the model, which is convenient while you are experimenting and risky once real work depends on the results. Keeping both in your codebase, alias for development and snapshot for production, costs nothing and removes a whole class of surprise.
The second habit worth keeping is a stored set of prompts you retest after every release. Each generation reads context slightly differently, so wording tuned for the previous model is the thing most likely to break quietly. A dozen prompts covering your real use cases will tell you within an hour whether an upgrade is safe. That is a cheaper answer than discovering the difference in production.
The third is to expect the interface to change more than the API does. Endpoints and request shapes have stayed stable across these releases, while the product side keeps gaining tools such as Sketch, Templates and pinned comments. If your work runs through the API, upgrades have been small. If it runs through ChatGPT, the new tools are the part worth relearning after each release.
Frequently asked questions
What is the newest GPT Image model?
GPT Image 2.5, released on 2026-09-08 as ChatGPT Images 2.5. In the API it appears as two models, `gpt-image-2.5-flare` for fast everyday generation and `gpt-image-2.5-sunburst` for editing precision. Both shipped with dated snapshots carrying the same release date, so you can pin the exact version rather than following the moving alias.
Is GPT Image 2 still available?
Yes. GPT Image 2 released on 2026-04-21 and remains a documented model with its own reference pages. It supports up to 16 reference images per edit request. Its token rates are identical to those of the GPT Image 2.5 models, which means cost is not a reason to stay on it or a reason to move off it.
Were mona-lisa-1 and luna-lisa-alpha official previews?
No. Their appearance in August 2026 is a community report and remains speculation. OpenAI has not confirmed what those evaluation-arena checkpoints were, and no release note names either of them. Articles that describe them as an official preview of GPT Image 2.5 are presenting an unverified rumour as fact, so treat their claims accordingly.
What was the hidden ChatGPT modal found on 2026-09-03?
A community report described a modal sitting behind a disabled feature flag in ChatGPT, reading that image creation got a major upgrade. This was a user finding, not an announcement, and OpenAI never commented on it. The release arrived days later, which is why the report is remembered, but it was never official at the time.
How often does OpenAI release a new image model?
Three generations shipped between December 2025 and September 2026, which averages out to a few months apart. That is a description of what happened rather than a schedule. OpenAI has not published a release cadence for image models, so no future date can be inferred from the gaps between the three releases listed above.
Do I need to update my code when a new model ships?
Not immediately, if you pinned a dated snapshot, since a snapshot keeps serving the version you tested. If you call the floating alias, your results can change when OpenAI updates the model underneath. Either way, retest your real prompts after a release. Wording tuned for an older generation is what usually breaks first.
Did GPT Image 2.5 change pricing?
No. Token rates for both Flare and Sunburst match GPT Image 2 exactly, covering text input, image input, their cached rates and image output. OpenAI has never published a per-image price for any of these models, because the token count depends on your prompt and your output size. Third-party hosts publish per-image prices of their own.