The News, Explained
OpenAI announced Images 2.6 on November 4. It is an updated model for generating and editing images from instructions or visual references. The company says it is available globally across all subscription tiers in ChatGPT and Codex on desktop, mobile, and web. App availability and developer controls in the API, which lets software call the models, describe different parts of the release. Source
GIF generation adds short moving images. API additions include markup to identify editing locations, SVG output for scalable graphics, and a complexity control for planning effort. Quality controls visual detail; complexity controls how much planning goes into a request. The API offers Flare for speed and Sunburst for greater detail and editing consistency. OpenAI claims 50% faster generation and says Flare at low quality can take as little as six seconds. That is a conditional vendor claim; comparing working times requires matching requests and settings. Source
The system card explains safety testing and limitations. It evaluated 3,077 adversarial requests intended to produce policy-violating images, separating safe outputs from unsafe outputs that moderation blocked or missed. These are challenging test cases, so the rates describe that evaluation rather than ordinary users’ exposure. Source
Unsafe outputs that were not blocked changed from 1.92% to 1.72% for Sunburst and from 1.49% to 2.14% for Flare against their corresponding baselines. Neither overall difference was statistically significant. Reading the direction of a percentage without that condition would overstate what the comparison establishes. Source
GIF safeguards face the additional difficulty of harmful meaning conveyed through motion. The card acknowledges remaining detection gaps. It also describes combining C2PA provenance metadata with invisible SynthID watermarks. These provide different signals about origin, while the card acknowledges that provenance has no single solution. Source
OYOPICK’s Take
More control over revisions is a welcome direction. If people can communicate both the location and planning needs of an edit, refining an image could become easier. We would assess that possibility through revision counts and preservation of intended details alongside first-generation speed. Demonstrating a saving in actual working time still requires a comparison of real tasks. Source
As moving images broaden creative possibilities, we hope review tools develop that consider both individual frames and their combined motion. The GIF detection limitations and evaluation conditions documented in the system card provide concrete points against which to assess that progress. Source