GPT-Live 系统卡：评测配置更正后部分安全指标出现回归

2. Model Data and Training

Like OpenAI’s other models, GPT-Live-1 and GPT-Live-1 mini 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.

Note that comparison values from previously launched models are from the latest versions of those models, so may vary slightly from values published at launch for those models.¹

3.1 Voice-Native Evaluations for Disallowed Content

August 4, 2026: We identified a configuration mismatch in the evaluation setup. We re-ran the full evaluation using the correct configuration, and the results reported in this section have been updated accordingly. We’ve added an appendix to show the original and corrected values side by side.

Appendix A: Original and corrected values for model safety evaluations

This section was added on August 4, 2026.

Compared with the previously reported results, the corrected production evaluation scores show statistically significant regressions on illicit behavior for GPT-Live-1 (0.97 to 0.95) and GPT-Live-1 mini (0.94 to 0.91). On the synthetic evaluations, GPT-Live-1 shows statistically significant regressions on sexual content (0.97 to 0.95) and gore (0.97 to 0.93), while GPT-Live-1 mini shows statistically significant regressions on illicit behavior (0.97 to 0.95) and gore (0.96 to 0.90). None of the regressions fall below our safety launch standards.