ChatGPT Images 2.6 System Card
ChatGPT Images 2.6 System Card
OpenAI released the ChatGPT Images 2.6 System Card, covering two image models, GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare. For the cross-frame risks of GIF generation, a safety reasoning model combines generation descriptions with frame grids for moderation; in a targeted evaluation of 310 annotated violating GIFs, the recall rate was 77.66%, and dynamically related violations remain difficult to detect in full. The Preparedness Framework evaluation adapts existing language model capability evaluations to the image modality, requiring the model to generate a reasoning draft and a final answer within the image and scoring only the final answer; neither model crossed the Bio High or Cyber High thresholds, but OpenAI followed the precautionary practice of Images 2.0 and 2.5, applying mitigations as for a high biological capability model, using a safety reasoning model to enforce biological risk policies on all inputs and outputs, and combining human review to monitor misuse, banning relevant accounts when sustained misuse patterns are detected.
The system card discloses that GIF violation moderation had a recall rate of 77.66% on 310 samples, and explains the precautionary consideration of applying safeguards as if for high biological capability even when the threshold was not crossed.