GPT Image 2.5: Flare vs Sunburst, and Why We Have Not Priced Them Yet

OpenAI did something unusual with GPT Image 2.5. Instead of one model replacing the last one, they shipped two, and made you choose.

Both are live on fal. We have verified the endpoints, the resolutions and the quality tiers. We have not put either into the Studio, and the reason is not quality. It is that neither model has a price we can quote you.

The Short Version

Flare is the everyday one. It is the direct successor to GPT Image 2 for general work, and OpenAI's headline claim is up to 50% lower latency than the model it replaces. If you are generating a batch of images and want them back quickly at good quality, this is the default.

Sunburst is the careful one. It is built for edits where exactly one thing should change and everything else must survive untouched. Swap the background behind a product while the product itself stays pixel-identical. Rewrite the labels on a diagram without redrawing the diagram. Change the headline on a campaign layout while the composition holds. It takes longer per image, deliberately.

That is the whole decision. Speed, or control over a constrained edit.

What They Share

Both variants run the same resolution ladder:

• 1024×768, 1024×1024, 1024×1536
• 1920×1080
• 2560×1440
• 3840×2160

And the same five quality tiers: low, medium, high, xhigh, max. Sunburst defaults to high.

Both also ship a text-to-image endpoint and a separate edit endpoint, which is what makes Sunburst's precision claim meaningful rather than marketing. The edit path is where a model either holds the untouched parts of an image steady or quietly redraws them, and that is the thing Sunburst is tuned to get right.

Which One To Reach For

Use Flare when the image is the deliverable and you want it now. Concept frames, social assets, batch work, anything where you will generate several options and pick one. The latency difference is the point.

Use Sunburst when the image already exists and you need a surgical change to it. Product photography where the item cannot shift. Anything with text in it that has to stay legible and correctly placed. Work where a second attempt costs you more than the extra generation time does.

If you are starting from a blank prompt rather than an existing image, Flare is almost always the right call. Sunburst's advantage lives in the edit path.

There is a sensible pattern that falls out of this, though it is inference from how the two models are documented rather than anything OpenAI prescribes: explore wide in Flare, pick your direction, then move the precision-sensitive refinement to Sunburst. You pay the slower model's time only on the one frame that earned it. That is the same iterate-cheap-then-commit shape that already works well for stills into video, and it is how we would use the pair.

The Pricing Problem

Here is where it gets awkward, and where we part company with most coverage of these models.

Both variants bill by tokens, not per image. The published rates are identical for each:

Token rates are not a price per image, and the gap between the two is not small. fal's own documentation says it plainly: longer prompts cost more, more complex requests cost more, and larger images cost more. The sample figures bear that out. Flare at 1024×1024 runs from roughly $0.0059 at low to $0.211 at max. That is a thirty-six-fold spread on one resolution, before prompt length enters into it.

And OpenAI has explicitly said the GPT Image 2 cost calculator does not estimate 2.5's token consumption, so you cannot derive a 2.5 quote from the older model's numbers either.

Why That Keeps Them Out Of The Studio

We price every model against the provider's published rate before switching it on. A model whose cost we cannot verify is a model we might be selling below cost, so it does not ship, however good the launch reel is.

That rule kept Seedance 2.5 out of the Studio for weeks while its endpoint existed and its rate did not. It is the same rule here, for a slightly different reason: the rate exists, but it is expressed in a unit that does not convert to a per-image price without knowing how many tokens a given generation will consume. Nobody publishes that, because it depends on what you ask for.

The honest position is that we could put a number on it and probably be right most of the time. We are not going to, because "probably right most of the time" is how you end up quoting five credits for a generation that costs you eleven.

What Happens Next

Two things unblock this. Either fal publishes a per-tier, per-resolution grid we can price against, or we run a controlled set of test generations at a few candidate tiers, including deliberately long and complex prompts, and price against the observed worst case with a buffer on top.

We will do the second if the first does not arrive. When it is priced, both variants go into the Studio with the cost on the button before you tap it, the same as every other model there.

One honest limit on this post. We have verified the endpoints, the resolutions, the tiers and the published token rates. We have not run a real brief through either variant. Nothing above is a quality verdict, and the Flare-versus-Sunburst guidance reflects what OpenAI and fal document about each model's design, not our own side-by-side test. When we have run one properly, this post gets the results.