What is GPT Image 2.5?

GPT Image 2.5 is the image model family behind ChatGPT Images 2.5, the image generation and editing system OpenAI released for ChatGPT and for its developer API. It makes pictures from a written description. It also edits pictures you upload. OpenAI shipped it to ChatGPT, ChatGPT Work and Codex users at every tier, on desktop, on mobile and on the web. The API exposes the same family as two named models you pick between on each request.

The system accepts text and images as input. It returns images only. It does not write captions, answer questions or hand back structured data. Everything you ask for arrives as a rendered picture. That single output type is worth planning around in an API integration, because your own code has to store and serve the file that comes back.

The gains over the previous generation sit in the rendering rather than in the interface. Lighting reads as more natural. Textures carry more detail. A face or an object taken from a reference photo survives an edit more often. Instruction following holds up better when you keep editing the same picture across several turns. Those four improvements are the ones OpenAI names, and they match what most people notice within a few prompts.

Inside ChatGPT the release also brought new ways to steer a generation. Sketch lets you type @Sketch and draw a rough layout that the model reads as a visual reference. Templates give you prebuilt formats such as posters and product photos. You can pin a comment onto one part of an image and ask for an edit only there. You can share a finished image together with the prompt that produced it. Every edit is kept in a version history, so you can walk an image back to an earlier state.

The API splits GPT Image 2.5 into two models. The model named `gpt-image-2.5-flare` is the default. It is the fast one, built for everyday generation, and it is the one to reach for first. The model named `gpt-image-2.5-sunburst` trades speed for editing precision. It runs slower by design. Send it the jobs where the exact placement of an element matters more than the wait does.

Speed is the clearest split between them. Flare answers markedly faster than the previous generation, and outside evaluations put it well ahead of GPT Image 2 on the same work. Sunburst is the slower of the new pair. The exact latency figures OpenAI published sit in the questions at the end of this page.

Safety behaviour differs slightly too. Both models refuse unsafe requests more often than Images 2.0 did. Sunburst has the lowest unsafe generation rate of the three. Flare sits between Sunburst and the older baseline. The published rates are close enough that they should not drive your model choice on their own.

Choosing between the two comes down to what the request is for. A first draft, a batch of variations, a thumbnail or a mood board belongs on Flare. A retouch where a logo has to stay legible, a product shot where a label must not shift, or a multi step edit that has to respect earlier instructions belongs on Sunburst. Most teams end up running Flare for volume and Sunburst for the handful of images that ship.

Price does not enter the decision. Both models charge the same token rates, so moving a request to Sunburst costs you time rather than money. Image output dominates the bill in almost every workload, and image output is priced identically across the pair.

OpenAI also publishes dated snapshots of both models. A snapshot pins the exact version you are calling. If you are building something where output has to stay consistent over months, call the snapshot rather than the floating alias. The alias moves when OpenAI updates the model. The snapshot does not move.

GPT Image 2.5 is the third named generation in this family. GPT-Image-1.5 came first. GPT Image 2 followed and shipped alongside ChatGPT Images 2.0. GPT Image 2.5 now sits on top of both. The dates for all three are on our release timeline page. Each generation replaced the one before it in practice while the older models stayed callable, so nothing broke for anyone who did not upgrade on release day.

The comparison that matters for most projects is GPT Image 2 against GPT Image 2.5. Token rates are the same for both, so moving up does not change your unit economics. Flare is faster than GPT Image 2 by every measure OpenAI and outside evaluators have published. Sunburst is the slower model of the new pair but the more precise editor. GPT Image 2 stays available, and it accepts a batch of reference images in a single edit request.

GPT-Image-1.5 is the oldest of the three and the weakest at the things this generation emphasises. If you are still calling it, the migration is mostly a change of model string. The endpoints are unchanged. The request shape is unchanged. Test your prompts again after switching, because the newer models read context differently and wording tuned for 1.5 will not always land the same way.

The naming trips people up more than the models do. ChatGPT Images 2.5 is the product inside ChatGPT. GPT Image 2.5 is the model family behind it, and in the API that family is `gpt-image-2.5-flare` and `gpt-image-2.5-sunburst`. Someone saying GPT Image 2.5 usually means one or the other. The distinction only matters when you read pricing, because token rates apply to API calls and not to the ChatGPT product.

You do not need a ChatGPT subscription to generate an image with these models. The API is the direct route. You create an OpenAI developer account, add credit, then call the generations endpoint with a model string and a prompt. You pay for the tokens you use and nothing else. That suits anyone who already writes code and wants the raw model with no interface in the way. Setup takes a few minutes and the first call is a single request with a model name and a sentence of description.

Third-party hosts are the other route. Providers such as fal.ai serve the same models behind their own API and their own per-image pricing. You accept a markup in exchange for knowing the price of an image before you generate it. The fal.ai figures in the table above show what that pricing looks like in practice.

This site sits in the same category. You bring a prompt, we run the model, and the image comes back in your browser without any API key setup. images25 is an independent site and is not affiliated with OpenAI. Every claim about model behaviour here comes from the OpenAI documentation linked at the bottom of the page.

The model behaves the same whichever route you take. A prompt that works in ChatGPT works through the API. The differences live in the wrapper: which quality settings are exposed, how reference images get attached, and whether a version history is kept for you. Those are product decisions made by whoever built the interface. They are not properties of the model itself.