Flare and Sunburst share rates, not the same bill

OpenAI has split GPT Images 2.5 into two models, Flare and Sunburst, and lists both at the same token rates: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. The launch, reported by Thenewstack, leaves the per-image cost comparison between the two unexplained.

At a glance

• The headline capability is targeted editing: a single element such as a product, a background or a line of copy can be changed while the surrounding scene stays intact.
• OpenAI positions Flare as the default option, saying it produces higher-quality images than GPT-Image-2 with 50% lower latency, while Sunburst trades speed for tighter control across edits.
• Developers cannot price the choice in advance: OpenAI publishes no token-consumption guidance for GPT Images 2.5 and says the GPT Image 2 calculator does not estimate it.

Identical rate cards make the two models look interchangeable on price, but rates are only half of a token bill; consumption is the other half, and that half is unpublished. For teams running image generation at volume, the practical effect is that the Flare-versus-Sunburst decision likely has to be made empirically, by generating the same assets through both and measuring. Cost modelling before a build appears to be off the table.

OpenAI puts Flare at 50% lower latency than GPT-Image-2

OpenAI calls Flare "the default choice for most applications" and says it delivers higher-quality images than GPT-Image-2 with 50% lower latency, a comparison drawn against the previous model rather than against Sunburst. The workloads listed for it are everyday ones: social content, rapid image prototyping, visual search and image generation.

Sunburst is positioned as the precision option. OpenAI describes it as "built for premium visual workflows that benefit from tighter control across edits", with longer generation times named as the trade-off. That trade-off appears only in the launch announcement; the model pages for Flare and Sunburst do not put a figure on the difference in generation time between them.

Both models carry the same $5, $8 and $30 per-million token rates

Image pricing is a function of token count, and the rate card is only one input into it. OpenAI does not publish an explicit method for estimating GPT Images 2.5 token consumption, so identical rates do not translate into comparable per-image costs for Flare and Sunburst.

Developers looking to OpenAI's existing GPT Image 2 cost calculator for an estimate will not get one for the new models. The company states the limitation itself, in the same place where it confirms that the token rates carry over unchanged from GPT Image 2:

Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption.

The rate parity extends backwards as well: the $5, $8 and $30 figures are the same ones GPT Image 2 carried, so the generation change brings no headline price increase. What it does not bring is a way to convert those rates into an expected bill per image.

Precision editing targets a product, background or line of copy

The core editing change is scope. GPT Images 2.5 can apply a change to a single element of a picture, such as a product, a background or a piece of copy, while preserving the rest of the scene, and it follows editing instructions more closely across multiple rounds of edits.

For API teams, OpenAI points to more reliable reference-led workflows: higher image fidelity is meant to keep each variation more closely tied to the original source. The company also says the model is "better at understanding complex visual instructions and translating them into coherent results", alongside intelligence and style improvements.

How much longer Sunburst takes is not stated anywhere OpenAI has published: not in the launch announcement, which mentions only "longer generation times", and not on the model pages for Flare or Sunburst. Nor is there a token-consumption estimate for either model, or word of an updated calculator. Until those appear, the comparison has to be made empirically.