Model profile
Claude Sonnet 4
Anthropicdeveloper
2025-05-22release date
#31 / 305overall rank
35eval lineages
Evidence summary
Claude Sonnet 4 has an estimated overall rank of #31; its 90% source-sensitivity interval is #24–#98. Its behavior-only rank is #39; company governance moves the combined estimate to #31. Published evidence spans 35 evals and 7 of 7 behavior components. Its strongest relative result is MASK (lying_probability_pct, #1 of 53); its weakest is Anthropic Claude Sonnet 4.5 System Card (harmful_request_safety, #2 of 2).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗claude-4-sonnet
- OpenRouter ↗anthropic/claude-sonnet-4
- System card ↗Family-level model document · Anthropic · first party
- Release source ↗direct research preview identity date
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #18 / 327 | ↓0.2912 | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #4 / 80 | ↑0.883 | Source ↗official | |
| Alignment Leaderboardcorrigibility | #7 / 24 | ↑4.398 | Source ↗official | |
| Alignment Leaderboardhonesty | #5 / 24 | ↑4.088 | Source ↗official | |
| Alignment Leaderboardnon_manipulation | #4 / 24 | ↑4.498 | Source ↗official | |
| Alignment Leaderboardrobustness | #10 / 24 | ↑3.813 | Source ↗official | |
| Alignment Leaderboardsafety | #4 / 24 | ↑4.606 | Source ↗official | |
| Alignment Leaderboardscheming | #4 / 24 | ↑4.422 | Source ↗official | |
| Anthropic Agentic Misalignment — blackmailmisaligned_action_rate_pct | #8 / 16 | ↓79 | Source ↗official | |
| Anthropic Agentic Misalignment — corporate espionagemisaligned_action_rate_pct | #3 / 16 | ↓17 | Source ↗official | |
| Anthropic Agentic Misalignment — lethal actionmisaligned_action_rate_pct | #3 / 10 | ↓22 | Source ↗official | |
| Anthropic Claude 4 System Cardagentic_coding_safety | #1 / 3 | ↑0.9 | Source ↗official | |
| Anthropic Claude 4 System Cardbenign_request_refusal | #2 / 3 | ↓0.0023 | Source ↗official | |
| Anthropic Claude 4 System Cardharmful_request_safety | #1 / 3 | ↑0.9899 | Source ↗official | |
| Anthropic Claude 4 System Cardstrongreject_jailbreak_success | #1 / 3 | ↓0.0266 | Source ↗official | |
| Anthropic Claude Sonnet 4.5 System Cardagentic_coding_safety | #2 / 2 | ↑0.893 | Source ↗official | |
| Anthropic Claude Sonnet 4.5 System Cardclaude_code_malicious_refusal | #2 / 2 | ↑0.8 | Source ↗official | |
| Anthropic Claude Sonnet 4.5 System Cardharmful_request_safety | #2 / 2 | ↑0.9822 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #51 / 106 | ↑0.295 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #31 / 44 | ↓86.6 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #42 / 50 | ↓75 | Source ↗official | |
| CAIS Risk Indexmask | #14 / 52 | ↓10.7 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #16 / 49 | ↓16 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #10 / 104 | ↓3.091 | Source ↗official | |
| Confabulationsconfabulation_rate | #3 / 52 | ↓3.96 | Source ↗official | |
| Constitutional Following — Anthropic Constitutionconstitutional_following_score | #6 / 7 | ↑85 | Source ↗official | |
| Emergent Collusionhigh_illegality_game_rate | #5 / 13 | ↓0.32 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #25 / 241 | ↑42.89 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #237 / 241 | ↑63 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #6 / 241 | ↑99.44 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #13 / 239 | ↑99.77 | Source ↗official | |
| FlagEval Safety and Valuesa1_qualified_rate | #1 / 18 | ↑87.01 | Source ↗official | |
| FlagEval Safety and Valuesa2_qualified_rate | #3 / 18 | ↑82.22 | Source ↗official | |
| FlagEval Safety and Valuesa3_qualified_rate | #2 / 18 | ↑89.41 | Source ↗official | |
| FlagEval Safety and Valuesa4_qualified_rate | #2 / 18 | ↑91.83 | Source ↗official | |
| FlagEval Safety and Valuesa5_qualified_rate | #18 / 18 | ↑49.88 | Source ↗official | |
| FORTRESSaverage_risk_score | #20 / 49 | ↓21.21 | Source ↗official | |
| FORTRESSover_refusal_score | #27 / 46 | ↓5.14 | Source ↗official | |
| HELM Safetyanthropic_red_team | #41 / 80 | ↑0.988 | Source ↗official | |
| HELM Safetybbq | #10 / 80 | ↑0.9725 | Source ↗official | |
| HELM Safetyharmbench | #13 / 80 | ↑0.9605 | Source ↗official | |
| HELM Safetysimple_safety_tests | #26 / 80 | ↑0.9975 | Source ↗official | |
| HELM Safetyxstest | #27 / 80 | ↑0.969 | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #44 / 54 | ↑24.49 | Source ↗official | |
| MASKlying_probability_pct | #1 / 53 | ↓7.7 | Source ↗official | |
| MonitoringBench Full-Trajectory Monitorfull_trajectory_catch_rate_at_1pct_fpr_percent | #4 / 13 | ↑42.1 | Source ↗official | |
| OpenAgentSafetyllm_judge_safety_vulnerable | #1 / 7 | ↓49.06 | Source ↗official | |
| OpenAgentSafetyrule_based_safety_vulnerable | #6 / 7 | ↓49.06 | Source ↗official | |
| OpenAgentSafetysuccessful_completion | #2 / 7 | ↑37.1 | Source ↗official | |
| PacifAIstp_score | #5 / 7 | ↑83.76 | Source ↗official | |
| PropensityBenchscore | #2 / 14 | ↓12.2 | Source ↗official | |
| RealityTest — Text AI-Identity Disclosuredisclosure_probability | #6 / 17 | ↑0.515 | Source ↗official | |
| Shelleducation_jsr | #1 / 14 | ↓0.28 | Source ↗official | |
| Shellfinance_jsr | #1 / 14 | ↓0.174 | Source ↗official | |
| Shellmanagement_jsr | #1 / 14 | ↓0.17 | Source ↗official | |
| Social Welfare Function Benchmarkfairness | #10 / 19 | ↑0.49 | Source ↗official | |
| SOSBenchbiology_pvr | #1 / 23 | ↓0.104 | Source ↗official | |
| SOSBenchchemistry_pvr | #6 / 23 | ↓0.21 | Source ↗official | |
| SOSBenchmedicine_pvr | #1 / 23 | ↓0.213 | Source ↗official | |
| SOSBenchpharmacology_pvr | #1 / 23 | ↓0.234 | Source ↗official | |
| SOSBenchphysics_pvr | #4 / 23 | ↓0.145 | Source ↗official | |
| SOSBenchpsychology_pvr | #2 / 23 | ↓0.123 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #84 / 105 | ↑6.48 | Source ↗official | |
| SpeciEvalland_animal_4ns | #43 / 105 | ↓4.47 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #20 / 105 | ↓4.5 | Source ↗official | |
| SpeciEvalspeciesism | #52 / 105 | ↓2 | Source ↗official | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #54 / 94 | ↑89.7 | Source ↗official | |
| Vigil Mental Health Safetyoverall_score | #8 / 23 | ↑53 | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.