Publish model system cards with capability evaluations
| Entity | Anthropic (Company) |
| Category | transparency |
| Date made | May 21, 2024 |
| Venue | Seoul AI Safety Summit |
| Status | kept |
Status Definition
The commitment has been fulfilled as stated. Evidence demonstrates complete implementation.
Commitment
Make public technical documentation of model capabilities, limitations, and safety evaluations for all deployed frontier models.
Evidence Timeline
Oct 10, 2026
Independent evaluation shows Claude 3.5 Sonnet maintains 94% refusal rate on clearly harmful requests, demonstrating safety measures documented in system card.
Sources:
- Refusal Behavior Evaluation (Sample)Aligned check — Aligned Super Intelligence
Oct 10, 2026
Status changed to kept
Anthropic published Claude 3.5 Sonnet system card including detailed capability evaluations and Responsible Scaling Policy (RSP) assessment.
Sources:
- Introducing Claude 3.5 Sonnet — Anthropic
Know More?
Have evidence about this commitment's status? Email [email protected] with sources.