The Pulse
OpenAI Details Safeguards for ChatGPT Images 2.6, Adding GIF Generation and Photo Markup
The ChatGPT Images 2.6 System Card was published on November 4, 2026. Its evaluations compare Sunburst and Flare with Images 2.5 and detail safeguards for image generation, GIFs and provenance.
AI.info Team · Written with AI · fact-checked ·
What ChatGPT Images 2.6 adds
The ChatGPT Images 2.6 System Card was published on November 4, 2026, describing the safety evaluations, mitigations and limitations for its updated image-generation models. The release adds GIF generation and photo markup in ChatGPT and the API, as well as SVG output support in the API. OpenAI says its updated Sunburst and Flare models improve editing consistency, detail preservation, and responses to sketches and image annotations. Flare is optimized for faster generation, while configurable reasoning effort gives users in Chat and developers in the API more control over generation and editing speed.
The card says ChatGPT Images 2.6's core safety stack is based on the same safety foundations as Images 2.5, with additional safeguards for GIF creation. OpenAI says the models were trained on diverse datasets that include publicly available internet information, information accessed through third parties, and material provided or generated by users, human trainers and researchers. Its data processing pipeline uses rigorous filtering to maintain data quality and mitigate potential risks, including advanced processes intended to reduce personal information in training data. Safety classifiers also help prevent or reduce harmful or sensitive content, including sexual content involving a minor.
Checks before and after generation
For Sunburst, OpenAI describes multiple image-specific safeguards. LLM-based policy checks assess requests before generation, and requests judged to violate policy are refused at that stage. A safety reasoning model then monitors text and image inputs, and blocks generation if it determines that any input violates policy. The model also checks the final image before it is shown to the user, blocking outputs it determines violate policy.
OpenAI says safety-focused image and text classifiers have been part of its image safety stack since the original GPT-4o image-generation launch. The company describes continuous improvements to its safety reasoning model, policies and monitoring, alongside synchronous CSAM classification for ChatGPT image generation. Its monitoring includes broader offline and online enforcement, with multiple evaluation stacks for higher-risk categories around minors.
How Sunburst and Flare compared with Images 2.5
OpenAI used an automated evaluation to measure selected production moderation checks against a shared set of 3,077 challenging requests designed to generate policy-violating images. The evaluation compared each Images 2.6 model with its corresponding Images 2.5 baseline using a common reference-grading procedure. OpenAI cautions that the adversarial test set does not represent how often violative requests or outputs arise in production, and that the results do not measure the complete deployed safety system.
In the overall results, Sunburst's unsafe-generation-not-blocked rate was 1.72%, compared with 1.92% for its Images 2.5 baseline. Flare's rate was 2.14%, compared with 1.49% for its baseline; OpenAI reports those as 66 and 46 requests, respectively, out of 3,077. Neither overall difference was statistically significant. The proportion of Flare outputs graded non-violative before moderation fell from 78.36% to 75.04%, a statistically significant decline; OpenAI says the largest contributions came from the Other and sexual prompt categories.
Of 20 net additional unsafe Flare outputs that were not blocked, 14 were in the Other category, which includes long-tail categories such as religious figures, fabricated or compromising depictions of real people, jailbreak attempts and image-based requests. OpenAI says no individual prompt category showed a statistically significant increase in violative outputs when safety moderation was applied. The card reports that, across the evaluated configurations, Sunburst produced fewer violative images after moderation than its Images 2.5 baseline.
GIF safeguards, capability tests and provenance
OpenAI identifies risks specific to GIFs, including sexualized movement and harmful meaning conveyed across frames. It says GIF safeguards are designed to detect and block policy-violating GIF content, and that some GIF rules block categories more broadly than rules for still images to account for inferred motion. The company tested those safeguards on a targeted set of adversarial GIFs labeled as policy-violating, while warning that motion-related violations remain challenging to detect and the safeguards may not catch all violations.
For its Preparedness Framework assessment, OpenAI adapted existing language-model capability evaluations to test image-generation capabilities in cybersecurity and biological-risk categories. It says neither GPT-Image-2.6 Sunburst nor GPT-Image-2.6 Flare crosses the Bio High or Cyber High threshold. OpenAI nevertheless applies mitigations appropriate for a model treated as having High biological capability, including an image-specific adaptation of its biological-risk policy to all ChatGPT Images 2.6 inputs and outputs.
For provenance, OpenAI says Images 2.6 maintains a commitment to C2PA metadata and incorporates Google DeepMind's SynthID watermarking through ChatGPT, Codex and the OpenAI API. The card says readers can check supported files for provenance signals with OpenAI's public verification tool.