GPT-Image-2.6: Faster image generation with Flare and Sunburst

OpenAI splits the work between its two new GPT-Image-2.6 API models by task. In the description of its launch video, OpenAI says to use Flare for a quick first pass on a poster or product shot, and Sunburst at medium or high quality when details need to stay consistent across edits. Low, medium and high are quality levels a developer can request. At the low setting, OpenAI says, Flare can produce an image in as little as six seconds; the description gives no time for Sunburst. OpenAI also says it is bringing GPT-Image-2.6 to Codex, its coding agent.

"Consistent across edits" describes a common way to use an image API: generate a picture, ask for a change, then ask for another, and hope the parts you did not mention stay put. OpenAI's example is changing a product's backdrop while keeping the product's shape and texture. Its model pages describe the pair the same way, Flare as "image generation and editing for fast iteration" and Sunburst as "image generation and editing for detailed visual work". Both pages list image inputs and OpenAI's image edit endpoint, so either model can take an existing picture and revise it. OpenAI does not say whether a Flare draft should be finished in Sunburst, or how much steadier Sunburst's edits are.

The headline number is thinner than it looks. OpenAI says the 2.6 versions of Flare and Sunburst generate images 50% faster than before, but the description does not name the earlier model, quality setting or kind of image behind that comparison, so a developer cannot turn it into an expected wait. The six seconds is a best case for Flare at low quality, not a typical time. The video description gives no prices. OpenAI's model pages, as of November 5, 2026, list identical per-token rates for both 2.6 models: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens, the same rates the 2.5 model pages show. Because billing is per token and neither 2.6 page says how many tokens an image uses, equal rates do not establish an equal cost per image.

Both model pages also point to markup-guided edits, SVG output and animated GIF output, which is what the description's suggestion to bring a mascot to life as a looping GIF draws on. Those features are explained in the context for the OpenAI Developers post announcing the same API release. The Flare and Sunburst pairing itself began with Images 2.5 in September, when OpenAI already advised Sunburst for editing precision and Flare for fast everyday generation.