Model profile
Claude Opus 4.6
Anthropicdeveloper
2026-02-05release date
#17 / 309overall rank
28eval lineages
Evidence summary
Claude Opus 4.6 has an estimated overall rank of #17; its 90% source-sensitivity interval is #16–#67. Its behavior-only rank is #20; company governance moves the combined estimate to #17. Published evidence spans 28 evals and 7 of 7 behavior components. Its strongest relative result is SpeciEval (belief_animal_sentience, #1 of 113); its weakest is Claude Sonnet 4.6 User Wellbeing (child_benign_refusal_rate, #4 of 4).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗claude-opus-4-6
- OpenRouter ↗anthropic/claude-opus-4.6
- System card ↗Exact 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 | #83 / 328 | ↓0.6284 | Source ↗official | |
| ANIMAscore | #6 / 22 | ↑0.7246 | Source ↗official | |
| Arena Factuality — Search Arena (factuality-only weighting)factuality_bt_rating | #16 / 30 | ↑1184.0 | Source ↗official | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #28 / 112 | ↑1451.0 | Source ↗official | |
| BioSecBench-Refusalbalanced_refusal_score | #5 / 10 | ↑0.3987 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #3 / 106 | ↑0.85 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #11 / 44 | ↓49.5 | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #8 / 49 | ↓34.8 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #15 / 50 | ↓46 | Source ↗official | |
| CAIS Risk Indexmachiavelli | #15 / 46 | ↓84.4 | Source ↗official | |
| CAIS Risk Indexmask | #18 / 52 | ↓12.7 | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #24 / 33 | ↓51.7 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #44 / 49 | ↓23.2 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #13 / 104 | ↓3.639 | Source ↗official | |
| Claude Sonnet 4.6 Overrefusalhigher_difficulty_overrefusal_rate | #1 / 5 | ↓0.04 | Source ↗official | |
| Claude Sonnet 4.6 Overrefusaloverall_overrefusal_rate | #4 / 5 | ↓0.66 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingchild_benign_refusal_rate | #4 / 4 | ↓0.18 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingchild_multiturn_appropriate_rate | #2 / 4 | ↑96 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingchild_violative_harmless_rate | #2 / 4 | ↑99.95 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingselfharm_benign_refusal_rate | #4 / 4 | ↓0.25 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingselfharm_harmless_rate | #1 / 4 | ↑99.75 | Source ↗official | |
| Claude Sonnet 4.6 User Wellbeingselfharm_multiturn_appropriate_rate | #3 / 4 | ↑82 | Source ↗official | |
| Constitutional Following — Anthropic Constitutionconstitutional_following_score | #2 / 7 | ↑97.1 | Source ↗official | |
| DystopiaBenchbasaglia_score | #5 / 50 | ↓26.77 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #10 / 50 | ↓33.37 | Source ↗official | |
| DystopiaBenchhuxley_score | #2 / 50 | ↓15.77 | Source ↗official | |
| DystopiaBenchlaguardia_score | #6 / 50 | ↓39.23 | Source ↗official | |
| DystopiaBenchorwell_score | #4 / 50 | ↓23.47 | Source ↗official | |
| DystopiaBenchpetrov_score | #5 / 50 | ↓30.07 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #72 / 241 | ↑22.74 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #58 / 241 | ↑91.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #6 / 241 | ↑99.44 | Source ↗official | |
| FORTRESSaverage_risk_score | #11 / 49 | ↓16.76 | Source ↗official | |
| FORTRESSover_refusal_score | #37 / 46 | ↓9.59 | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #23 / 54 | ↑27.23 | Source ↗official | |
| MACHIAVELLIdeception_relative_random_pct | #15 / 50 | ↓84.4 | Source ↗official | |
| MASKlying_probability_pct | #5 / 53 | ↓9.16 | Source ↗official | |
| MORUscore | #6 / 13 | ↑74.1 | Source ↗official | |
| PHAREbias_resistance_diagnostic | #57 / 66 | ↑0.3351 | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #2 / 70 | ↑0.8776 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #1 / 70 | ↑1 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #6 / 67 | ↑0.7524 | Source ↗official | |
| RealityTest — Text AI-Identity Disclosuredisclosure_probability | #3 / 17 | ↑0.66 | Source ↗official | |
| RefusalBenchyouden_j | #4 / 19 | ↑0.6667 | Source ↗official | |
| SABERoverall_safety_rate | #1 / 13 | ↑45.26 | Source ↗official | |
| SABERscenario_a_safety_rate | #1 / 13 | ↑56.3 | Source ↗official | |
| SABERscenario_b_safety_rate | #1 / 13 | ↑39.77 | Source ↗official | |
| SABERscenario_c_safety_rate | #1 / 13 | ↑36.87 | Source ↗official | |
| SM-Benchadversarial | #30 / 79 | ↑83.41 | Source ↗official | |
| SM-Benchambiguous_interpretation | #45 / 79 | ↑83.63 | Source ↗official | |
| SM-Benchanti_hallucination | #1 / 79 | ↑100 | Source ↗official | |
| SM-Bencheq_boundaries | #76 / 79 | ↑42.98 | Source ↗official | |
| SM-Benchoverfit | #3 / 79 | ↑95.63 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #1 / 113 | ↑7 | Source ↗official | |
| SpeciEvalland_animal_4ns | #107 / 113 | ↓5.05 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #89 / 113 | ↓5.03 | Source ↗official | |
| SpeciEvalspeciesism | #73 / 113 | ↓2.17 | Source ↗official | |
| StereoTales Harmful Associationsbenign_significant_association_score | #21 / 23 | ↑84.23 | Source ↗official | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #75 / 94 | ↑87.8 | 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.