Claude Sonnet 5.5 System Card Released

What happened

Anthropic released the system card for Claude Sonnet 5.5, a document describing model evaluations and safeguards. The source also references content involving multiple companies (Anthropic / Meta / xAI), and the system card is an official technical document for the model.

Key details

• The document states that evaluations were run on the "final released snapshot" of the model; different snapshots are taken at various points during training, and unless otherwise noted, all evaluations in the card used the final released snapshot.
• Models are compared on the same harness, so scores may differ slightly from those in previously published system cards.
• The evaluations cite GDPval-AA v2.1, an independent evaluation framework developed by Artificial Analysis, which uses 220 tasks from OpenAI's GDPval gold database spanning 44 occupations across nine major industries; tasks mirror actual professional work.
• The document discusses two autonomy threat models: one concerning systems with capacity for autonomous, goal-directed operation and subterfuge, which could plausibly (if directed toward this goal, deliberately or inadvertently) carry out sabotage, leading to "irreversibly and substantially higher odds of a later global catastrophe"; the other concerning risks from automated R&D by AI systems that can fully automate or dramatically accelerate the work of large, top-tier human teams.
• Anthropic Limited is the provider of Anthropic's general-purpose AI models in the European Economic Area.

Why it matters

A system card is an official first-hand source on a model's capabilities and safety evaluations. Independent benchmarks like GDPval-AA, built on real professional tasks, are closer to economically valuable work than traditional benchmarks. The document's treatment of autonomy threat models—particularly sabotage risks and automated R&D—reflects a formal framework for assessing high-risk scenarios at frontier labs.

Who it affects

• AI practitioners and researchers: the evaluation framework and snapshot mechanism are relevant for interpreting model performance.
• Safety and alignment communities: the threat model descriptions merit attention.
• Users in the European Economic Area: the document concerns the local service provider entity.

that every model is compared on the same harness. As a result, the scores may differ slightly from those in system cards published before that change. 8.14.3 GDPval-AA GDPval-AA v2.1, developed by Artificial Analysis, is an independent evaluation framework that tests AI models on economically valuable, real-world professional tasks. The benchmark uses 220 tasks from OpenAI’s GDPval gold database,20 spanning 44 occupations across nine major industries. Tasks mirror actual professional work [...] capacity for autonomous, goal-directed operation and subterfuge, such that it is plausible these systems could (if directed toward this goal, either deliberately or inadvertently) carry out sabotage, leading to irreversibly and substantially higher odds of a later global catastrophe. ●​ Autonomy threat model 2: risks from automated R&D in key domains. This threat model concerns AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human [...] Limited is the provider of Anthropic’s general-purpose AI models in the European Economic Area. To contact Anthropic, visit our Support page. 1.4 Model evaluations Different “snapshots” of our models are taken at various points during the training process. Except where otherwise noted, the evaluations in this system card were run on the final released snapshot of Claude Sonnet 5.5. 1.5 Safeguards The safeguards we have applied to Claude Sonnet 5.5 for general access usage are as follows: ●​ For

The original belongs to 官方博客 · Anthropic / Meta / xAI and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.