news

OpenAI ships Images 2.6 with a 50% faster claim and no stated baseline

The November 4 launch gives developers two API models, GPT-Image-2.6 Flare and Sunburst, one speed percentage with no measurement conditions, and a six-second figure that only applies at low quality.

By Daniel Ochoa · 5 November 2026

OpenAI's Images 2.6 launch on November 4 puts exactly one performance number in front of developers, 50% faster image generation, and does not say faster than which model, at what quality setting, at what output resolution, or on what kind of prompt. The company's own post states that "Images 2.6 generates images 50% faster than before" and leaves every variable behind that comparison unspecified.

The release itself is broad. Images 2.6 is live in ChatGPT and Codex across all subscription tiers on web, desktop and mobile, and two new models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, are available through the API. On the consumer side the headline additions are animated GIF output, where a prompt or an existing image can be turned into a loop with adjustable speed and per-frame edits, and saved reference images that keep a person, pet or character available for reuse without re-uploading. For developers, OpenAI added image markup, so an annotated image can be sent to the API to mark where an edit should land, SVG export, and a complexity parameter that adjusts how much planning the model does before it renders.

Six seconds is a low-quality number, and the 50% has no named baseline

The split between the two API models is the familiar one OpenAI used for the 2.5 generation. Flare is the latency tier and Sunburst is the detail tier, with OpenAI saying Sunburst holds details more consistent from one edit to the next when quality is set to medium or high. The company's line that "Flare generates images in as little as six seconds" is tied to the low quality setting, which makes it a floor rather than a typical generation time, and the 50% figure is never attached to a named predecessor, a resolution or a measured workload. That is a vendor claim with no reproducible conditions, the same shape of claim we have seen when image model launches are scored entirely in-house.

The rate card is in the docs, not in the announcement

OpenAI's launch post carries no prices. The model pages do. Both gpt-image-2.6-flare and gpt-image-2.6-sunburst list image output at $30.00 per 1M tokens, image input at $8.00 per 1M with cached input at $2.00, and text input at $5.00 per 1M with cached input at $1.25. The two tiers are priced identically, which means the choice between Flare and Sunburst is a latency and consistency decision rather than a cost one, and the real cost variable for anyone running an edit chain is how many retries the pipeline needs.

The ranking claim came from a third-party board, not from OpenAI's own numbers

OpenAI's developer account pointed people to Arena's text-to-image and image editing results rather than publishing any evaluation of its own, and DesignArena posted that GPT Image 2.6 Sunburst took the top spot on Image Arena ahead of its predecessor, 2.5 Sunburst. No margin was attached to that claim in the launch thread, which matters because image arena leads are routinely narrow enough that confidence intervals overlap. OpenAI also puts volume across ChatGPT and the API at more than 4 billion images a week, up from 3 billion in September.

What is verifiable today: the models exist, they are documented, the rate card is published, and the GIF and reference-image features are shipping to all tiers. What is not verifiable today: the only number OpenAI chose to lead with.

Sources: OpenAI, OpenAI API docs, GPT-Image-2.6 Flare, OpenAI API docs, GPT-Image-2.6 Sunburst, OpenAI on X.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.