OpenAI Releases Images 2.6 With GIFs and API Complexity: ‘50% Faster’ Lacks a Baseline, Six Seconds Is a Low-Quality Best Case; In an Offline Test of Selected Production Moderation Checks, Flare’s Violative Rate Rose From 1.49% to 2.14% Without Statistical Significance

On November 4, 2026, OpenAI introduced Images 2.6 in ChatGPT, Codex and the API, with a system card published alongside. The release does three things: it makes image generation faster, it adds GIF output, and it gives developers more control over editing rounds. Two models carry the name in the API, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. The speed and volume figures below come from OpenAI's announcement, and the safety figures come from its system card.

On speed, OpenAI says Images 2.6 generates images 50% faster than before. The post does not say what "before" means, so we cannot say which earlier model, quality setting or prompt set the 50% is measured against. The more concrete claim is that Flare, with quality set to low, can generate images in as little as six seconds. That is a best case at the lowest quality, not a typical time. Flare is described as the faster option and as better at layout for posters and product imagery. Sunburst is the choice for more detail and for consistency across edits, when quality is set to medium or high. OpenAI also says people now create 4 billion images per week across ChatGPT, Codex and the API, more than half a billion a day. That is OpenAI's own usage figure.

The new feature is GIF generation. Marketers and creators can animate a brand mascot, make reaction loops, or sketch a storyboard before video production. OpenAI shows a sprite-sheet prompt that asks for six frames in a three-by-two grid, and says even a rough hand-drawn sketch can become an animated character. Several partners appear as examples: OpenArt, which says one prompt is enough to speak the native language of social media, Town, a startup that used animated characters as a typing indicator in a work assistant, Figma Weave, Picsart and others. These are early-access customer statements, not independent tests.

For developers, three things arrive in the API. Image markup, which already existed in ChatGPT, lets a team point to exactly where a change should happen. SVG export makes outputs usable on websites, apps and print. A new complexity parameter sets how much planning the model does for a request. OpenAI recommends low complexity for most editing tasks, paired with the image quality the workflow needs, and medium or high complexity for scenes with many elements or constraints. Note that quality and complexity are separate controls: quality sets visual detail and fidelity, while complexity guides how much goes into the scene. The post links to a pricing page instead of stating prices, so this article gives no price.

Availability is stated plainly. Images 2.6 is available globally today in ChatGPT and Codex across all subscription tiers, on desktop, mobile and web. Both API models are available now. The post does not list regional limits for the API.

The system card is where the caveats are. OpenAI compared each new model with its GPT-Image-2.5 counterpart on a shared set of 3,077 challenging requests built to produce policy-violating images. These results are not representative of production traffic. This offline evaluation covers selected production moderation checks, not the full deployed safety system; the saved violative outputs are not outputs delivered to users. Sunburst produced fewer violative images after moderation than its baseline. For Flare the rate rose from 1.49% (46 requests) to 2.14% (66 requests), an increase of 0.65 percentage points that OpenAI says is not statistically significant. Before moderation, the share of Flare outputs graded non-violative fell from 78.36% to 75.04%, and OpenAI reports that this decline was statistically significant, with most of the change in the "other" and sexual categories. Of 20 net additional unsafe outputs that moderation did not block, 14 were in the "other" category.

GIFs add risks that single images do not, such as sexualized movement and harmful meaning spread across frames. OpenAI applies GIF-specific rules that block some categories more broadly than for still images. On a separate targeted set of GIFs already labeled violative, whose sample size is not disclosed, moderation recall was 77.7% overall, with 77.8% for sexual content and 77.5% for self-harm, and higher figures for hate and abuse (90.9%), violence and gore (89.7%), extremism (92.9%) and violent activities (96.6%). OpenAI states that motion-related violations remain hard to detect and that the safeguards may not catch all of them. The card also reports that neither model crosses the biological or cybersecurity High thresholds, while OpenAI still applies its stricter biological-risk policy as a precaution. For provenance, it keeps C2PA metadata and adds Google DeepMind's SynthID watermarking across ChatGPT, Codex and the API.

Our reading: the announcement is a product update with real new capabilities, and the numbers that can be checked are the system card's. The 50% speed gain has no stated baseline, the six-second time is a low-quality best case, and the usage count is self-reported. The GIF safeguards detect roughly three in four GIFs already labeled violative in the adversarial set in OpenAI's test set, so teams that publish generated GIFs should keep their own review step.