Model profile
Claude 3 Sonnet
Evidence summary
Safety. Claude 3 Sonnet has an estimated Safety rank of #234; its 90% source-sensitivity interval is #110–#251. Its behavior-only rank is #251; company governance moves the combined estimate to #234. Published Safety evidence spans 15 eval lineages and 7 of 7 components. Its strongest relative result is HELM Safety (simple_safety_tests, #1 of 80); its weakest is Claude 3.5 Sonnet model-card safety and alignment evaluations (incorrect_refusals_xstest, #4 of 4).
Freedom. Claude 3 Sonnet has an estimated Freedom rank of #643; its 90% source-sensitivity interval is #500–#638. Published Freedom evidence spans 11 eval lineages and 1 of 1 components. Its strongest relative result is Claude 3.5 Sonnet model-card safety and alignment evaluations (human_feedback_harmlessness_win_rate_pct, #1 of 5); its weakest is Claude 3 model-card adversarial human-preference evaluations (multimodal_harmful_response_rank, #2 of 2).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗claude-3-sonnet
- OpenRouter ↗anthropic/claude-3-sonnet
- System card ↗Family-level model document · Anthropic · first party
- Release source ↗direct research preview identity date
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AgentDojotargeted_attack_success_rate | #12 / 15 | ↓0.2671 | Source ↗official | |
| AgentDojoutility_under_attack | #12 / 15 | ↑0.3323 | Source ↗official | |
| AgentHarmharm_score | #7 / 12 | ↓20.7 | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #13 / 80 | ↑0.847 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #22 / 104 | ↓7.827 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationscorrect_refusals_wildchat_rank | #2 / 5 | ↓2 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsdiscrimination_rank | #2 / 5 | ↓2 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationshuman_feedback_harmlessness_win_rate_pct | #1 / 4 | ↑87 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationshuman_feedback_honesty_win_rate_pct | #1 / 4 | ↑69 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsincorrect_refusals_wildchat_rank | #2 / 5 | ↓2 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsincorrect_refusals_xstest_rank | #4 / 5 | ↓4 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsmultimodal_hallucination_rank | #1 / 2 | ↓1 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsmultimodal_harmful_response_rank | #1 / 2 | ↓1 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationscorrect_refusals_wildchat | #2 / 4 | ↑93.6 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationshuman_feedback_harmlessness_win_rate_pct | #5 / 5 | ↑47 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationshuman_feedback_honesty_win_rate_pct | #4 / 5 | ↑38 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationsincorrect_refusals_wildchat | #2 / 4 | ↓8.8 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationsincorrect_refusals_xstest | #4 / 4 | ↓36.6 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #81 / 248 | ↑21.71 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #9 / 248 | ↑96.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #40 / 248 | ↑93.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #16 / 246 | ↑99.64 | Source ↗official | |
| HELM Safetyanthropic_red_team | #7 / 80 | ↑0.998 | Source ↗official | |
| HELM Safetybbq | #55 / 80 | ↑0.9 | Source ↗official | |
| HELM Safetyharmbench | #15 / 80 | ↑0.958 | Source ↗official | |
| HELM Safetysimple_safety_tests | #1 / 80 | ↑1 | Source ↗official | |
| HELM Safetyxstest | #74 / 80 | ↑0.858 | Source ↗official | |
| Large-scale Moral Machine experiment on LLMshuman_choice_distance | #30 / 39 | ↓1.249 | Source ↗official | |
| OR-Benchover_refusal_rate | #22 / 25 | ↓94.5 | Source ↗official | |
| OR-Benchtoxic_acceptance_rate | #2 / 25 | ↓0.3 | Source ↗official | |
| SORRY-Benchavg | #9 / 51 | ↓0.11 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AIRBench 2024 Safety Scenariossafety_scenarios | #68 / 80 | ↓0.847 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #83 / 104 | ↑7.827 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationscorrect_refusals_wildchat_rank | #4 / 5 | ↑2 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationshuman_feedback_harmlessness_win_rate_pct | #3 / 4 | ↓87 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsincorrect_refusals_wildchat_rank | #2 / 5 | ↓2 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsincorrect_refusals_xstest_rank | #4 / 5 | ↓4 | Source ↗official | |
| Claude 3 model-card adversarial human-preference evaluationsmultimodal_harmful_response_rank | #2 / 2 | ↑1 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationscorrect_refusals_wildchat | #3 / 4 | ↓93.6 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationshuman_feedback_harmlessness_win_rate_pct | #1 / 5 | ↓47 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationsincorrect_refusals_wildchat | #2 / 4 | ↓8.8 | Source ↗official | |
| Claude 3.5 Sonnet model-card safety and alignment evaluationsincorrect_refusals_xstest | #4 / 4 | ↓36.6 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #238 / 248 | ↓96.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #208 / 248 | ↓93.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #228 / 246 | ↓99.64 | Source ↗official | |
| HELM Safetyanthropic_red_team | #72 / 80 | ↓0.998 | Source ↗official | |
| HELM Safetyharmbench | #65 / 80 | ↓0.958 | Source ↗official | |
| HELM Safetysimple_safety_tests | #58 / 80 | ↓1 | Source ↗official | |
| HELM Safetyxstest | #74 / 80 | ↑0.858 | Source ↗official | |
| OR-Benchover_refusal_rate | #22 / 25 | ↓94.5 | Source ↗official | |
| OR-Benchtoxic_acceptance_rate | #21 / 25 | ↑0.3 | Source ↗official | |
| SORRY-Benchavg | #42 / 51 | ↑0.11 | Source ↗official |