OpenAI 发布 GPT-Image-2.6（Sunburst 与 Flare）系统卡

2. Model Data and Training

Like OpenAI’s other models, GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare were trained on diverse datasets, including information that is publicly available on the internet, information that we partner with third parties to access, and information that our users or human trainers and researchers provide or generate. Our data processing pipeline includes rigorous filtering to maintain data quality and mitigate potential risks. We use advanced data filtering processes to reduce personal information from training data. We also employ safety classifiers to help prevent or reduce the use of harmful or sensitive content, including explicit materials such as sexual content involving a minor.

3.2 Safety and Policy Evaluations

3.2.2 GIF Risks, Safeguards, and Evaluations

GIF generation introduces risks that may not be apparent in individual still images, including sexualized movement and harmful meaning conveyed across frames. To address these risks, the safety reasoning model evaluates the generated frame grid alongside the generation caption. GIF-specific rules also block some content categories more broadly than in still images to account for inferred motion.

On a targeted evaluation of 310 GIFs labeled policy-violating, moderation using the generation caption and frame grid achieved 77.66% recall. These results measure performance on the evaluation set, not the frequency of harmful outputs in production. Motion-related violations remain challenging to detect, and the safeguards may not catch all violations.

3.3 Preparedness Framework: Image-Specific Capability Assessment and Safeguards

OpenAI’s Preparedness Framework is designed to track and prepare for models with frontier capabilities that could create new risks of severe harm in three tracked categories: Biological and Chemical, Cybersecurity, and AI Self-Improvement. Because image models are unable to create and execute code in a way that would enable AI Self-Improvement, we do not have evidence that GPT-Image-2.6 Sunburst nor GPT-Image-2.6 Flare pose meaningful risk in this category.

For the Cyber and Biological categories, we adapted existing language-model capability evaluations to assess image-generation capabilities. For each task, we instructed the model to generate an image containing both a visible reasoning scratchpad and a final answer. We then graded only the final answer rendered in the image, using the evaluation’s existing grading criteria. This allowed us to test the model on the same underlying cyber-capability and biological-capability questions while eliciting and evaluating its response through its native image modality.

The results indicate that neither GPT-Image-2.6 Sunburst nor GPT-Image-2.6 Flare cross the Bio High threshold or the Cyber High threshold.

Although neither model crosses these thresholds, we follow the precautionary approach used for images 2.0 and 2.5 and apply mitigations appropriate for a model treated as High biological capability. These mitigations include an image-specific adaptation of our existing biological-risk safety policy, which we apply to all ChatGPT Images 2.6 inputs and outputs using our safety reasoning model. The general input and output blocking mechanisms are described in Section 3.1.

3.3.1 Biological-risk monitoring and enforcement

As with GPT-Images 2.0 and 2.5, we are enabling the same policy enforcement checks for biorisk that we do with our text-based models, largely outlined in this article. We use advanced reasoning models with high biological capabilities to detect biological misuse, combining our automated systems with human reviewers to monitor and enforce our policies. We have integrated the analysis from our image-based safety reasoning model into the signals we use to detect ongoing misuse. In some cases where we detect ongoing patterns of misuse we suspend the offending accounts.