Claude Mythos Preview

Memorization screens flag a subset of problems. Excluding flagged problems, Mythos Preview's margin over Opus 4.6 holds. • Self-reported

GPQA Diamond. • Self-reported

Scores higher than Opus 4.6 while using 4.9× fewer tokens. • Self-reported

CharXiv Reasoning with tools. Without tools: 86.1%. Opus 4.6: 78.9% (with tools), 61.5% (without tools). • Self-reported

100% pass@1. Benchmark saturated. • Self-reported

Cybersecurity vulnerability reproduction. • Self-reported

LAB-Bench FigQA with tools. Opus 4.6: 75.1% (with tools). • Self-reported

GraphWalks BFS 256K–1M. Opus 4.6: 38.7%, GPT-5.4: 21.4%. • Self-reported

With tools. Without tools: 56.8%. Anthropic notes Mythos may show some level of memorization at low effort. • Self-reported

Multilingual Q&A. Opus 4.6: 91.1%, Gemini 3.1 Pro: 92.6–93.6%. • Self-reported

• Self-reported

Memorization screens flag a subset of problems. Excluding flagged problems, Mythos Preview's margin over Opus 4.6 holds. • Self-reported

Internal implementation. Scores not directly comparable to public leaderboard scores. Opus 4.6: 27.1% on same implementation. • Self-reported

Memorization screens flag a subset of problems. Excluding flagged problems, Mythos Preview's margin over Opus 4.6 holds. • Self-reported

Terminus-2 harness with adaptive thinking at maximum effort, 1M token total task budget. 1× guaranteed / 3× ceiling resource allocation, averaged over 5 attempts per task. 92.1% with 4-hour timeout limits and Terminal-Bench 2.1 updates. • Self-reported

USAMO 2026 math proofs. Opus 4.6: 42.3%, GPT-5.4: 95.2%, Gemini 3.1 Pro: 74.4%. • Self-reported