GPT-6.1 Sol 使用与 GPT-6 Astra 相同的数据和训练方法
Addendum to GPT-6 Astra System Card: GPT-6.1 Sol
GPT-6.1 Sol 使用与 GPT-6 Astra 相同的数据和训练方法，在生物学和化学领域被评估为高能力，未跨越临界阈值。在网络安全评估中表现优于所有先前生产模型，但在合成和半合成智能体环境中较 GPT-5.6 Sol 略有下降，模型拒绝仍是安全边界的一层。
2. Model Data and Training
GPT-6.1 Sol uses the same types of data and training as GPT-6 Astra, described in the GPT-6 Astra card.
For previously launched models, the values published at launch reflect the versions evaluated at that time. The comparison values for previously launched models that are shown here may reflect later versions of those models, and may vary from the values published at launch.1
9.1.1 Biological and Chemical Capabilities
We are treating GPT-6.1 Sol as High capability in the Biological and Chemical domain. Below, we report results for GPT-6.1 Sol on our High capability evaluations as well as results for our Critical capability evaluations. GPT-6.1 Sol’s reported results did not cross the indicative Critical thresholds.
For a full description of these evaluations, please see the Biological and Chemical Capabilities section in the GPT-6 Astra card.
9.2.1 Model Safety Training and Evaluation
9.2.1.1 Biological and Chemical Safety Training and Evaluation
In biology refusal evaluations, GPT-6.1 Sol performs comparably to GPT-6 Sol in the severe and dual-use categories, while refusing fewer benign prompts. These results reflect model responses alone, without our full production safeguards.
9.2.1.2 Cybersecurity Safety Training and Evaluation
On the cybersecurity safety evaluations, GPT-6.1 Sol outperforms all our previous models in production-chat evaluations. Compared to GPT-5.6 Sol, GPT-6.1 Sol shows modest regressions in synthetic and semi-synthetic agentic environments. Model refusal remains one layer of our safety stack, alongside additional safeguards that enforce the safety boundary through defense in depth.
来源: OpenAI Deployment Safety · deploymentsafety.openai.com