Anthropic outlines 10 recommendations for stronger AI safety evaluations

Anthropic published safety evaluation recommendations and disclosed that Claude Fable 5.1 scored 60% via the API versus 94% on claude.ai on a self-harm safety test.

On November 13, 2026, Anthropic published its Building strong safety evaluations research post, laying out 10 recommendations for designing, running, and reporting the safety evaluations that gate model releases, and disclosing a notable scoring gap for Claude Fable 5.1.

The 10 recommendations

The post lists 10 recommendations for designing, running, and reporting safety evaluations that gate model releases
It draws an analogy between weak AI safety evaluations and psychology's replication crisis
Recommendations include construct validity, preregistration, human validation of automated graders, and multi-turn simulations for sensitive requests
Anthropic calls for multidisciplinary evaluation teams, real-time monitoring during evaluations, and transparent reporting of results

A scoring gap in Claude Fable 5.1

Anthropic disclosed that Claude Fable 5.1 scored 60% on a multi-turn suicide and self-harm safety evaluation when accessed via the API, compared with 94% on claude.ai. Anthropic attributed the difference to claude.ai-specific system-prompt safety instructions that are absent by default via the API.

What it means for people building with AI

The disclosed scoring gap is a concrete reminder for developers building on the Claude API that safety behavior tuned into the consumer claude.ai product does not automatically carry over to raw API usage. Builders who rely on the API need to add their own safeguards rather than assume claude.ai-level protections apply by default.

Sources

Building strong safety evaluations (Anthropic, November 13, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.