On mvanemmerik.com, Nov 16: NFL Week 10 chalk talk: Detroit dazzles, Cincinnati takes the North
vanemmerik.aiClaude, ChatGPT and Grok news, twice a day

Saturday, November 14, 2026 · Claude & Anthropic

Claude · Safety · happened Nov 13, 2026

Anthropic says single-turn safety tests are saturated, with recent Claude models scoring 94% to 98%

In a research post, Anthropic says its single-turn harmlessness tests now mostly serve as regression checks, since every recent model scores 94% to 98%, and it is leaning on simulated multi-turn conversations graded by another model. It reports that results for Claude Fable 5.1 differed by up to 34 percentage points between the API and claude.ai, and that multi-turn confidence intervals can exceed plus or minus 10 points. It also says it hardened its test environments and paused many evals that give models internet access.

Why it matters

It shows how much a model's measured safety can depend on test design and product setup, not just the model.

Sources

AnthropicClaude

Also on Nov 14, 2026