Skip to content

Anthropic

Claude Mythos Preview

Never on general release: only Project Glasswing partners could use it, for defensive cybersecurity work. An early version has the highest METR time horizon published, about 17.4 hours; METR calls results above 16 hours unreliable.

Announced
Access
Limited preview

Benchmark scores

TestScoreBest on the trackerHow it was measured
METR time horizontask length an AI agent finishes half the timeNo longer updated17 hTop scoreIndependentearly snapshot, METR Horizon v1.1 run 'claude_mythos_preview_early_inspect'; 95% CI 508.9–3304.3 min

Compare every model on the same tests →

Also on these charts

Sources