OpenAI releases "ChatGPT Images 2.6" with GIFs and reference images, adds 2 "50% faster" models to API

GIFs and reference images, plus 2 new API models

OpenAI released "ChatGPT Images 2.6," a new version of its image generation, on Nov. 4 (Nov. 5 Japan time). It added a feature that creates GIFs from text prompts or photos and a reference image feature for reusing saved images. For its developer API, it introduced new models "GPT-Image-2.6 Flare" and "GPT-Image-2.6 Sunburst," saying image generation is 50% faster. The launch follows up on a leak in which the 2.6 name and GIF preparations were found in documentation and code from late October, with OpenAI now making a formal announcement.

In ChatGPT, users can change the playback speed of GIFs they create and fix them frame by frame. If users save images of themselves, pets or characters, they can use them in new images without uploading them again.

According to OpenAI, more than 4 billion images a week are created in ChatGPT and the API, up from 3 billion in September. The figures are the company's own, and it did not give a breakdown for GIFs.

Rolling out to all ChatGPT and Codex users

Images 2.6 has started rolling out to all users of ChatGPT and Codex on the web, desktop and mobile. The rollout is gradual, and according to OrcaRouter, some accounts could not use it on the first day. Generation limits by plan are not stated in the documents that could be checked.

50% faster and as little as 6 seconds are OpenAI's claims

OpenAI's developer account posted that Flare and Sunburst generate images 50% faster than before. The documents that could be checked do not say what the figure is compared with or under what conditions it was measured. The model description pages also have no speed figures.

According to OpenAI, Flare creates images in as little as 6 seconds when quality is set to low. Sunburst, with quality set to medium or high, preserves detail even through repeated edits, the company said. The 6 seconds figure applies only to Flare's low-quality setting.

Third-party figures do not match OpenAI's claims. The evaluation site Arena posted that 2.6 is nearly twice as fast as 2.5. According to OrcaRouter, in a comparison by another evaluation site, Design Arena, Flare was about 2.6 times as fast as the previous generation and Sunburst about 1.15 times. No third-party measurement verifying "50%" and "6 seconds" under the same conditions has been found.

According to the documentation, processing complex prompts can take up to 2 minutes. A new "complexity" setting, which determines how much planning is done before generation, defaults to the highest level, and higher levels can take longer.

First on Design Arena, second on Arena

OpenAI said Sunburst ranked first on the evaluation site Design Arena's "Image Arena." Design Arena also posted on X on the day of the announcement that Sunburst had taken first place, overtaking its predecessor, 2.5 Sunburst.

The scores, however, cannot be verified. The Design Arena leaderboard checked on Nov. 5 was the Nov. 4 update from before the announcement, with no row for 2.6, and 2.5 Sunburst led with 1,416 points. OrcaRouter also described it as a ranking without publicly verifiable scores.

A post by Arena, another evaluation site where users compare and vote on outputs from models whose names are hidden, shows a different ranking. In the text-to-image category, Sunburst (medium quality) was second with 1,409 points, and Flare (medium quality) was fourth with 1,384 points. In the style category "Cartoon, Anime & Fantasy," Sunburst ranked first.

As of Nov. 5, 2.6 was not on Artificial Analysis' leaderboard. The previous-generation 2.5 Sunburst led with 1,224 points, and OpenAI held the top three spots. On the editing leaderboard, 2.5 Sunburst's 1,199 points was followed by Microsoft's "MAI-Image-2.7-Preview" at 1,196. The scores are Elo ratings calculated from voting wins and losses, with no maximum.

API pricing same as 2.5, 2 supported endpoints

API pricing is the same for Flare and Sunburst and unchanged from the previous generation, 2.5. Per million tokens, input costs $5 for text and $8 for images, and output costs $30. There is no fixed price per image, and costs vary with quality and size.

The cost gap with competitors is large. In Artificial Analysis' estimates, 1,000 images at 1024×1024 cost $210.7 with 2.5 Sunburst (max). Google's "Nano Banana 2.1" costs $33.6, "MAI-Image-2.6" $38.9 and Meta's "Muse Image" $10. Estimates for 2.6 are not yet listed.

In addition to complexity, the API supports "markup," which lets users draw marks on an image to show where to make layered fixes, as well as SVG and GIF output. SVG and GIF are limited to new generations and cannot be used for editing. Markup cannot be used together with the existing method of specifying the area to fix by painting over it. On the model description pages, the supported endpoints are image generation and editing, and the Responses API and batch processing are not included.

Leaked features become official

Markup, found in the API documentation on Oct. 28, was released in the changelog as a feature supported by 2.5 and 2.6. GIF output, found in Codex code on Nov. 3, also became a product. It has not been confirmed whether the limit of 2 to 16 frames in the code is the same in the released version.

aurora, which appeared in the API list on Oct. 18, had been seen as a candidate for the next "GPT Image 3," but the name released this time was 2.6. The relationship between aurora and 2.6 has not been confirmed.