OpenAI splits image generation in two: Flare for speed, Sunburst for precision

OpenAI has refreshed its image models and, on the developer side, now ships two models instead of one. GPT-Image-2.5 Flare is the default for most applications: in the company's words it delivers higher-quality images than GPT-Image-2 at 50% lower latency. GPT-Image-2.5 Sunburst is built for detailed creative work and trades longer generation times for tighter control across edits.

The split is the story on its own. Opening two endpoints under one version number, one for speed and one for precision, is an admission that a single model does not serve every job. A team producing campaign creative and a catalogue turning out hundreds of product shots do not want the same model.

Scale

The figure OpenAI gives: more than 3 billion images a week are created across ChatGPT Images and the GPT-Image models in the API. At that volume a 50% cut in latency is a cost line, not a convenience.

On the ChatGPT side

ChatGPT Images 2.5 shipped today to ChatGPT, ChatGPT Work and Codex users, across all tiers, on desktop, mobile and web. Image generation latency is down by up to 50% compared with Images 2.0.

Three interface features come with it. Sketch lets you draw inside ChatGPT and use the drawing as a visual guide; you open it by typing "@Sketch". Templates remove the blank canvas for common formats such as posters and product photos. You can also place comments directly on an image to steer a focused edit, and when you share an image you can include the prompt that produced it.

The real claim: knowing what not to change

The capability being sold here is not generation but restraint. Change the colour of a jacket while the subject stays in position. Replace a background without disturbing the product shot. Revise copy without reworking the composition. OpenAI also highlights multi-turn editing consistency in long conversations: earlier changes hold, and quality does not degrade as edits accumulate. Complex layouts and transparent backgrounds are named as improvements in this release.

What is missing: the price

The announcement gives no pricing figures and points to a pricing page instead. Read that alongside the picture we described on 4 September, when Microsoft conceded second place on image quality and attacked on price with "best price-per-Elo in the world". OpenAI's answer in this announcement is speed rather than quality, and the only number offered is a latency improvement measured against its own previous model, with no competitor comparison.

On safety, OpenAI says prompt and image checks remain in place and that it continues to use C2PA metadata and invisible watermarking on generated images.

Source: OpenAI announcement.