Frontier Red Team

How does a model perform on this Anthropic-defined Frontier Red Team evaluation?

Aggregate capability measurements from Anthropic's publisher-defined intelligence and simulated weapons-development evaluations. Higher is better. Cell-member classification F1 on the easy tier.

This is an Anthropic-defined evaluation rather than a common benchmark leaderboard. The task set, harness, model variant, tool access, reasoning setting, and safety constraints materially affect interpretation; compare only within this reporting context.

Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.78577; source interval 0.753319–0.818222; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.805056; source interval 0.782432–0.827679; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.803937; source interval 0.777644–0.83023; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.795413; source interval 0.76412–0.826706; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.805502; source interval 0.772156–0.838848; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.807631; source interval 0.773428–0.841835; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.819891; source interval 0.795897–0.843885; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.797919; source interval 0.765779–0.830058; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained. Anthropic's embedded identity-classification artifact reports cell-member classification F1 by difficulty tier; this row retains exact means, source intervals, and n=18. The artifact's upper-bound caveat is preserved in the context note. Identity classification · easy; exact aggregate mean 0.806478; source interval 0.777111–0.835844; n=18. Opus 4.5 and Opus 4.6 used a fixed 12k-token thinking budget where shown; all other models used adaptive thinking. Raw task material is not retained.