Anthropic: Single-Turn Safety Tests Show 94% to 98% Harmless Rates Across Recent Models

Anthropic published recommendations for more reliable AI safety evaluations on November 13, 2026, reporting aggregate harmless-response rates of 94% to 98% for every recent model on its single-turn API tests, according to a summary from Unite.AI.

Key points

• Saturated single-turn tests: Anthropic reports that every recent model scores between 94% and 98% in aggregate on its single-turn API tests.
• Remaining use: Because the tests are saturated, they are useful mainly as regression checks on known failure modes.
• Multi-turn evaluations: These offer more variation than single-turn tests and can capture gradual changes across a conversation.
• Definition: The guidance describes a safety evaluation as a measure of a model's propensity to cause or contribute to harm.

Details

The harmless-response figures are Anthropic's own results on its own single-turn API tests. The company reports one aggregate range covering every recent model, not a per-model breakdown. Unite.AI's summary does not name the individual models or give their scores.

Anthropic's position is that high scores on these tests leave little room to tell models apart. Their remaining value is catching regressions on failure modes that are already known. Multi-turn evaluations, in Anthropic's description, show more variation between models and can reflect how behavior shifts gradually over the course of a conversation, which single-turn prompts cannot.

What this means

When recent models all land between 94% and 98% on the same test, the test no longer separates one model from another. Anthropic's guidance treats these single-turn tests as a check that known failure modes have not returned. For questions about how a model behaves over an extended exchange, Anthropic points to multi-turn evaluations. Teams comparing models on safety should read a single-turn harmless-response rate in that context.