| Human Pathogen Capabilities Test v2 (HPCT v2) | 74.6% | | #3 of 71 | Official leaderboard |
| CyberBench v1.1 (Vals) | 77.9% | | #3 of 43 | Official leaderboard |
| TxBench-Antibody-Discovery | 63.7% | | #2 of 28 | Official leaderboard |
| TxBench-Preclinical-Pharmacology | 64.7% | | #2 of 28 | Official leaderboard |
| CyScenarioBench | 100% | | #1 of 14 | Model card |
| BioSecBench-Surveillance | 59.8% | | #2 of 27 | Official leaderboard |
| SRE-Bench | 44.3% | | #3 of 33 | Official leaderboard |
| CheatBench | 8.60% | | #1 of 10 | Official leaderboard |
| DrugDiscoveryBench | 65.4% | | #2 of 19 | Official leaderboard |
| CoT-Control | 83.7% | | #1 of 8 | Model card |
| GDM Situational Awareness | 90.9% | | #1 of 8 | Model card |
| EpiBench | 28.9% | | #4 of 27 | Official leaderboard |
| VariantBench | 50.3% | | #4 of 27 | Official leaderboard |
| OncologyBench | 33.3% | | #5 of 30 | Official leaderboard |
| GDM Stealth | 100% | | #1 of 6 | Model card |
| LinuxArena | 45.5% | | #1 of 6 | Model card |
| BioSecBench-Refusal | 48.7% | | #4 of 22 | Official leaderboard |
| scBench-Long | 28.4% | | #5 of 27 | Official leaderboard |
| TxBench-Oligo-Discovery | 51% | | #6 of 28 | Official leaderboard |
| BioSecBench-Function | 44.4% | | #7 of 27 | Official leaderboard |
| PCT (Polarized Contrastive Pairs) | 38.7% | | #12 of 46 | Official leaderboard |
| ExploitBench | 87% | | #5 of 19 | Model card |
| CWE-bench | 68% | | #4 of 12 | Official leaderboard |
| MetagenomicsBench | 42.3% | | #10 of 26 | Official leaderboard |
| Gray Swan IPI Arena benchmark | 0.70% | | #4 of 10 | Model card |
| Chem / Bio Dynamic Jailbreak Eval | 38% | | #1 of 2 | Model card |
| Cyber Adaptive Jailbreak Eval | 20% | | #1 of 2 | Model card |
| Cyber Conversational Evals | 0.40% | | #1 of 2 | Model card |
| Cyber Static Jailbreak Eval | 1.00% | | #1 of 2 | Model card |
| Gray Swan Bio Eval | 4.00% | | #1 of 2 | Model card |
| Gray Swan Cyber Eval | 10% | | #1 of 2 | Model card |
| Real-world vulnerability discovery | 86% | | #1 of 2 | Model card |
| SpatialBench-Long | 28.6% | | #15 of 27 | Official leaderboard |
| DistressBench | 68.6% | | #29 of 42 | Official leaderboard |
| CBRN Conversational Evals | 2.60% | | #2 of 2 | Model card |
| CBRN Science Eval | 6.10% | | #2 of 2 | Model card |
| Agent harness misuse evaluation | 91.4% | | #1 of 1 | Model card |
| Agentic coding honeypot monitor evasion | 47% | | #1 of 1 | Model card |
| Agentic task gaming and overreach evals | 0.49 | | #1 of 1 | Model card |
| Bio Static Jailbreak Eval | 0% | | #1 of 1 | Model card |
| Biological static hazard evaluation | 0.13% | | #1 of 1 | Model card |
| Child Safety Action Testing | 100% | | #1 of 1 | Model card |
| Child Sexual Abuse and Exploitation Safety | 100% | | #1 of 1 | Model card |
| Coding agent prompt injection robustness | 83% | | #1 of 1 | Model card |
| Computer use manual red teaming | 100% | | #1 of 1 | Model card |
| Computer use Prohibited Action Evaluation | 99.4% | | #1 of 1 | Model card |
| Computer use prompt injection robustness | 92% | | #1 of 1 | Model card |
| Contextual security evaluation | 27% | | #1 of 1 | Model card |
| Cyber agentic safeguard evaluation | 100% | | #1 of 1 | Model card |
| Dangerous Content Safety | 96.6% | | #1 of 1 | Model card |
| Delusional Reinforcement evaluation | 94% | | #1 of 1 | Model card |
| Emotional Attachment evaluation | 76% | | #1 of 1 | Model card |
| Google DeepMind scheming honeypot suite | 0% | | #1 of 1 | Model card |
| Gram (Gauging Realistic Agentic Misbehavior) | 0% | | #1 of 1 | Model card |
| Grounding robustness (RAG prompt injection) | 99% | | #1 of 1 | Model card |
| Harmful manipulation strategic planning red teaming | 98.2% | | #1 of 1 | Model card |
| Hate & Harassment Safety | 92.7% | | #1 of 1 | Model card |
| HelpBench | 76.5% | | #1 of 1 | Model card |
| Impossible Tasks Evaluation | 93.8% | | #1 of 1 | Model card |
| Instrumental convergence evals | 0% | | #1 of 1 | Model card |
| Internal codebase monitor evasion eval | 76.4% | | #1 of 1 | Model card |
| Internal vulnerability research | 72% | | #1 of 1 | Model card |
| Neutrality evaluation | 93% | | #1 of 1 | Model card |
| Overstepping boundaries on ambiguous and high-risk tasks | 55.9% | | #1 of 1 | Model card |
| Personal-device assistant prompt injection eval | 98% | | #1 of 1 | Model card |
| Prompt injection, general tool use | 91% | | #1 of 1 | Model card |
| Prompt injection, turn injections in tool use | 86% | | #1 of 1 | Model card |
| RL rollout audit for beyond-episode drives | 0% | | #1 of 1 | Model card |
| Sexually Explicit Content Safety | 97.6% | | #1 of 1 | Model card |
| Stealth and Situational Awareness evaluations | 93.3% | | #1 of 1 | Model card |
| Suicide & Self-Harm evaluation | 96% | | #1 of 1 | Model card |
Google DeepMind