GPT Image 2.5

What it is

GPT Image 2.5 is the current head of OpenAI’s image line, and it arrived as two models rather than one. The company’s API changelog records the release on 8 September 2026 in a single line: “Released GPT Image 2.5 Sunburst and GPT Image 2.5 Flare for image generation and editing”, reachable through the Image API and through the image generation tool in the Responses API. Each variant carries a default snapshot stamped with that same day, gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08, which fixes the date on OpenAI’s own record.

The pair sits directly above GPT Image 2, the model that led the public boards from April through to the end of August 2026. On the 15 September 2026 reading of the arena.ai boards, OpenAI holds the top three seats on text-to-image and on image edit at once, and all three are its own: Sunburst, Flare, then GPT Image 2.

The two variants

The split between them is a split of purpose. OpenAI points Sunburst at workflows where editing precision comes first, and describes Flare as built for “fast, high-quality everyday image generation”. So the decision a team faces is about the shape of the job: a retouching or composition pipeline where each instruction has to land exactly, or a high-volume feed where speed and consistent quality carry the day.

Underneath, the two present the same interface. Both accept text and image as input and return an image, both are exposed on the same two API surfaces, and both appear on the pricing table at identical rates. That symmetry is useful in practice, because a team can route work between them on the strength of the task rather than rebuilding a pipeline around whichever one it picked first.

What the API gives you

The release brings two new quality settings. The full ladder on the image generation guide now reads low, medium, high, xhigh, max and auto, with xhigh and max added for this generation and named as such in the changelog entry. The top settings buy more compute per image, which is the lever a team pulls when a single frame has to be right rather than merely serviceable.

Output sizes are unchanged in shape. The guide recommends 1024x1024 for square work, 1536x1024 for landscape and 1024x1536 for portrait, and accepts custom dimensions written as a WIDTHxHEIGHT string, 1536x864 among the examples it gives.

Editing runs through the image edits endpoint, which covers three distinct jobs: changing an existing image, generating a new image using other images as a reference, and masking part of a frame so that only the selected area is regenerated. OpenAI draws one distinction between the two variants, on editing precision against everyday speed, and leaves the choice of endpoint job to the developer.

What it costs

Both variants bill at “GPT Image 2 token rates”, and OpenAI’s pricing table bears that out row for row. Text input runs at $5 per million tokens, cached text input at $1.25, image input at $8 per million with cached image input at $2, and image output at $30 per million. The model they sit above carries the same three headline numbers on the same page.

That is the commercially interesting part of this release. A step up in board position usually arrives with a step up in price, and teams plan around the gap. Here the budget written for GPT Image 2 carries across to either variant without amendment, which turns adoption into a change of model string rather than a procurement exercise.

Where it stands on the boards

Both arena.ai boards were dated 7 September when read on 15 September 2026, a week after release. These arenas rank models by putting their outputs against each other in blind human comparisons and converting the results into an Elo score, the rating method used in chess, so a position reflects accumulated human preference rather than a single benchmark run.

On text-to-image, Sunburst sits first at 1421 and Flare second at 1399, both marked preliminary, with GPT Image 2 (medium) third at 1381 and Microsoft’s MAI-Image-2.6 fourth at 1331. On image edit the gap is wider: Sunburst first at 1520 and Flare second at 1491, again marked preliminary, GPT Image 2 third at 1461, Grok Imagine Image 2.0 fourth at 1439, MAI-Image-2.6 fifth at 1434, and Google’s Nano Banana Pro ninth at 1390.

The preliminary label is doing real work on those top two rows. It means the boards are still gathering votes, and a preliminary score carries a wider band of uncertainty than a settled one, so the figures above are dated readings that will move. What is already firm is the ordering within OpenAI’s own line, where GPT Image 2.5 outranks GPT Image 2 on both boards at once.

How to access it

Access runs through the OpenAI API, under the model IDs gpt-image-2.5-sunburst and gpt-image-2.5-flare, on the Image API or the Responses API image generation tool. For a team already holding OpenAI credentials, that is an existing billing relationship and an existing integration, with the model string as the only variable to change.

Readers weighing the alternatives can start at our image models hub, which sets each system against the others, and the GPT Image 2 page covers the model this pair now sits above.

What to watch

The open question is where the preliminary scores settle. Sunburst and Flare entered the boards at the top and have held it through their first week, and the votes that will confirm or move those numbers are still being cast. The second thing to follow is the response from the rest of the field: Microsoft’s MAI-Image-2.6 and SpaceXAI’s Grok Imagine Image 2.0 trade fourth and fifth places beneath OpenAI’s three, close enough that a single release from either reorders the middle of both boards, while Google’s Nano Banana Pro sits ninth on the edit board.

Related

• GPT Image 2, the model this pair sits above on both arena.ai boards
• Nano Banana Pro and Nano Banana 2, Google’s line
• Grok Imagine Image 2, fourth on image edit
• Muse Image, Meta’s entry on the same board