Flare and Sunburst share rates, not the same bill

OpenAI’s GPT Images 2.5 reaches the API as two models, Flare and Sunburst. Both carry the same published token rates, and OpenAI gives no way to work out which one costs more per image or how much slower the precise one runs, The New Stack reported on September 10. The launch is mostly about editing: changing one part of an image without disturbing the rest.

OpenAI calls Flare “the default choice for most applications.” It is aimed at everyday image work: social content, rapid prototyping, visual search and general generation. OpenAI says Flare produces higher-quality images than GPT-Image-2 with 50% lower latency, which makes it an upgrade on the previous model on both counts it publishes.

Sunburst is the precise option, “built for premium visual workflows that benefit from tighter control across edits.” The obvious uses are production-ready campaigns and product imagery, where a stray change to a logo or a label means redoing the asset. That precision comes with “longer generation times,” in OpenAI’s words. The launch announcement and the model pages for Flare and Sunburst give no figure for how much longer.

What the editing upgrade covers

The main change is a precision editing ability. You can change a single element of a picture, such as a product, a background or a block of copy, and the rest of the scene should stay as it was. OpenAI also says the model now follows editing instructions more closely across several rounds of edits. In a production pipeline that matters, because a few small revisions should no longer mean regenerating the whole asset and checking it again from scratch.

Reference-led workflows also get higher image fidelity, so each variation stays closer to the source image it was built from. OpenAI says the model is “better at understanding complex visual instructions and translating them into coherent results,” and promises improvements in intelligence and style. The goal is for more images to come out right on the first try, before any edit is needed.

OpenAI has made this pitch before. GPT Image 1.5, released December 16, 2025, was already sold on precise edits that keep details intact, and it added face and logo preservation through the input_fidelity parameter. Version 2.5 makes the same promise again, now split across a fast tier and a precise one.

Same rates, unknown bill

Flare and Sunburst list identical token rates: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. Image generation is billed by tokens. Equal rates only mean equal bills if both models spend the same number of tokens on the same image, and OpenAI has published no guidance on estimating GPT-Image-2.5 token consumption for either model.

The existing cost calculator won’t fill the gap. OpenAI’s own note on it reads: “Token rates match GPT Image 2. The GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption.”

That leaves the comparison to the developer. Published pricing can’t tell you what the same kind of image costs on Flare versus Sunburst before you start, and the latency gap between them has no number either. The only way to find out whether Sunburst’s tighter control is worth the extra generation time, and possibly extra spend, is to run the same jobs through both models and compare token usage and wall-clock time on your own workload.