GPT Image 2.5 API

What is GPT Image 2.5?#

GPT Image 2.5 is OpenAI's image generation and editing model, released September 8, 2026 and not yet callable on Unifically. We will open it here once the integration is live, and every number on this page comes from the release and gets re-checked against the callable API on day one.

It arrives as two models rather than one. gpt-image-2.5-flare is the default: better pictures than GPT Image 2 at half the latency, which is what most production traffic wants. gpt-image-2.5-sunburst is slower and steadier, built for work that goes through many rounds of edits without drifting. Both take text and images in and return images, through the same generations and edits endpoints as GPT Image 2.

The gains over GPT Image 2 are in detail, lighting, and how well a subject survives an edit. In ChatGPT the same model powers Images 2.5, with sketch input and templates on top.

Key features of GPT Image 2.5#

• Two models, one API shape. Flare for speed and volume, Sunburst for edit precision. Same endpoints, same parameters, so switching between them is a one-word change.
• Half the latency of GPT Image 2. OpenAI measures Flare at 50% lower latency than GPT Image 2 while returning better images.
• New xhigh and max quality settings. Two steps above high, available only on the 2.5 models. Earlier GPT Image models stop at high.
• Subjects survive the edit. Faces, products, and distinctive details carry from a reference photo into a new setting, style, or composition instead of being redrawn.
• Edits that stay put over many turns. Change one element and the rest of the frame holds. Earlier edits stay intact as new ones stack on top, so a long chain does not degrade.
• Up to 4K with transparent backgrounds. Custom sizes to 3840 pixels on the long edge and 8.29 megapixels total, plus true transparency in png or webp.

Best for#

Use cases#

Build a product photo pipeline where one reference shot turns into a full campaign set: same bottle, new backgrounds, lighting, and crops. Wire an in-app editor that takes plain instructions like "swap the copy on the label" and changes only that, over as many rounds as a designer wants. Generate marketing collateral straight to a transparent png so it drops onto any background without a cutout step. And for anything with real information in it, like a slide, a diagram, or a menu, the layout and text hold together well enough to use, which was the weak point of every earlier image model.

Limitations#

Complex prompts can take up to two minutes. Text placement and clarity are better than GPT Image 2 but still miss on dense typography. Keeping a recurring character or a brand element identical across separate generations is not guaranteed, which matters for series work. Resolutions above 2560x1440 are marked experimental. Masked edits are guidance, not a precise selection: the model reads the mask as a hint and may not follow its exact shape, and with several input images the mask applies to the first one. The Arena numbers below rest on a few thousand votes each against hundreds of thousands for settled models, so those ranks are early and will move.

Flare vs Sunburst#

Sunburst wins on quality, Flare wins on speed, and the gap is smaller than the naming suggests. On Arena as of September 9, 2026, Sunburst leads Flare by 22 Elo on text-to-image (1421 to 1399) and 29 on image-edit (1520 to 1491). Both sit above GPT Image 2. Latency is where they split properly: Flare is built to return in half the time GPT Image 2 takes, and Sunburst is explicitly the slower of the two. So Flare is the right default for anything user-facing or high-volume, and Sunburst earns its extra seconds on final assets and long edit chains where a drifted detail costs more than the wait.

Availability#

OpenAI released ChatGPT Images 2.5 and both API models on September 8, 2026. In ChatGPT it reached every plan on desktop, mobile, and web the same day. Unifically pricing, variants, and exact callable parameters will appear on this page once the model is live here. Until then, GPT Image 2 is the current OpenAI image model on our API.

What GPT Image 2.5 can do#

First and second on both Arena image boards

As of September 9, 2026, Sunburst holds #1 on Arena text-to-image at 1421 Elo and #1 on image-edit at 1520 Elo, with Flare at #2 on both. GPT Image 2 is third on each, ahead of MAI Image 2.6, Grok Imagine Image 2.0, SeeDream 5.0 Pro, and Nano Banana Pro. OpenAI now holds the top three rows of both boards. The two 2.5 rows carry a few thousand votes each, so treat the exact Elo as early.

An edit that changes one thing and leaves the rest alone

Turn one generates a gray tabby cat hugging an otter in an orange scarf. Turn two is three words: "Now make it look realistic." The pose, the scarf, the hug, and the framing all survive; only the rendering changes. That is the multi-turn behavior developers were missing in GPT Image 2.

Four reference products composed into one shot

Body lotion, soap, an incense kit, and a bath bomb go in as four separate reference images. The model builds a single photorealistic gift basket that contains all four, adds the ribbon and the hand-lettered label the prompt asked for, and keeps each product recognizable. Useful for catalog and bundle imagery where the items have to be the real ones.

A face that stays the same face

A printed studio portrait goes in; the child comes back in an ivory tuxedo with black lapels and a bow tie. The pose, the blue backdrop, and the face are untouched. Subject preservation across a wardrobe or setting change is the single biggest gain over GPT Image 2, and the one that decides whether reference-led product work is worth building on.