Claude Mythos 5.1

The trusted-access twin of Claude Fable 5.1, released the same day (September 1, 2026): same underlying model, safeguards specifically designed to support cybersecurity and life-sciences work. Availability is narrow — US organisations and individuals in Anthropic's Life Sciences Verification Program, with a Cyber Verification Program described as coming rather than available — so no public per-token rate is recorded. Anthropic reports it designed very high-affinity protein binders with a ~50% hit rate across 12 targets (10-15% is typical) validated experimentally by two external organisations, and wrote custom GPU kernels that sped seven open-source deep-learning models up by as much as 2.5x with identical outputs. ONLY ONE BENCHMARK ROW IS RECORDED HERE. Anthropic's system card reports a single combined 'Fable 5.1 / Mythos 5.1' column, so almost every score is shared with Fable 5.1 and is filed there; duplicating it under two slugs would double-count the same measurement in the pairwise rating. Terminal-Bench 4.0 is the one row where the safeguard difference is broken out (60.9 unsafeguarded vs 55.8), and the launch post's accuracy-vs-cost chart attributes the whole Fable/Mythos gap to tasks where the earlier, less precise cyber safeguards intervened. That leaves this entry deliberately below the ranking-coverage floor.