| BioSecBench-Function | 53.5% | | #2 of 27 | Official leaderboard |
| scBench-Long | 31.8% | | #2 of 27 | Official leaderboard |
| VariantBench | 51.7% | | #2 of 27 | Official leaderboard |
| Prompt injection robustness against adaptive attackers, browser use (Anthropic) | 0% | | #1 of 13 | Model card |
| OncologyBench | 42.7% | | #3 of 30 | Official leaderboard |
| TxBench-Preclinical-Pharmacology | 64.3% | | #3 of 28 | Official leaderboard |
| BioSecBench-Surveillance | 54.6% | | #3 of 27 | Official leaderboard |
| EpiBench | 30.2% | | #3 of 27 | Official leaderboard |
| MetagenomicsBench | 60% | | #3 of 26 | Official leaderboard |
| DeepsecBench | 37.5 | | #4 of 28 | Official leaderboard |
| TxBench-Antibody-Discovery | 57.3% | | #4 of 28 | Official leaderboard |
| TxBench-Oligo-Discovery | 51.9% | | #5 of 28 | Official leaderboard |
| SHADE-Arena | 22.3% | | #3 of 16 | Model card |
| SRE-Bench | 30.1% | | #7 of 33 | Official leaderboard |
| SpatialBench-Long | 36.5% | | #6 of 27 | Official leaderboard |
| Single-turn benign request evaluations (Anthropic) | 0.02% | | #5 of 18 | Model card |
| DistressBench | 74.6% | | #12 of 42 | Official leaderboard |
| Virology Capabilities Test (VCT) | 58% | | #8 of 28 | Model card |
| CyScenarioBench | 46.1% | | #4 of 14 | Model card |
| Prompt injection robustness against adaptive attackers, computer use (Anthropic) | 0.07% | | #4 of 13 | Model card |
| Anthropic automated behavioral audit | 2.87 | | #2 of 6 | Model card |
| Binary Exploitation Benchmark (Anthropic, OSS-Fuzz based) | 85.8% | | #2 of 6 | Model card |
| LinuxArena | 36.2% | | #2 of 6 | Model card |
| Silent copying of answers (Anthropic) | 8.00% | | #1 of 3 | Model card |
| ExploitGym | 26.6% | | #5 of 14 | Model card |
| ExploitBench | 80% | | #7 of 19 | Model card |
| Anthropic political even-handedness eval | 97.9% | | #5 of 13 | Model card |
| WellnessBench | 57% | | #5 of 12 | Official leaderboard |
| MASK | 94.5% | | #9 of 21 | Model card |
| Self-preference bias (Anthropic) | 0.16 | | #2 of 4 | Model card |
| Gray Swan IPI Arena benchmark | 0.40% | | #6 of 10 | Model card |
| Malicious use of Claude Code (Anthropic) | 85.2% | | #10 of 16 | Model card |
| Anthropic undermining liberal democracy (behavioral audit metric) | 1.21 | | #4 of 6 | Model card |
| Concealment of changes (Anthropic) | 96.2% | | #2 of 3 | Model card |
| Sandbox escape evaluation (Anthropic) | 4.60% | | #2 of 3 | Model card |
| CyberBench v1.1 (Vals) | 59.6% | | #32 of 43 | Official leaderboard |
| BBQ (Bias Benchmark for QA) | 80% | | #18 of 24 | Model card |
| Prompt injection robustness against adaptive attackers, coding (Anthropic) | 3.01% | | #10 of 13 | Model card |
| Single-turn harmful request evaluations (Anthropic) | 95.6% | | #16 of 18 | Model card |
| Malicious computer use (Anthropic) | 79.5% | | #13 of 13 | Model card |
| Rewind-attacker cyber adversarial robustness (Anthropic) | 21% | | #5 of 5 | Model card |
| Cyber harm coverage (Anthropic) | 99.4% | | #4 of 4 | Model card |