Model profile
Claude 3 Haiku
Evidence summary
Claude 3 Haiku has an estimated overall rank of #161; its 90% source-sensitivity interval is #68–#197. Its behavior-only rank is #174; company governance moves the combined estimate to #161. Published evidence spans 17 evals and 7 of 7 behavior components. Its strongest relative result is HELM Safety (anthropic_red_team, #1 of 80); its weakest is HELM Safety (bbq, #77 of 80).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗claude-3-haiku
- OpenRouter ↗anthropic/claude-3-haiku
- 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 | Better | Source |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #134 / 311 | 0.7825 | ↓ lower | Source ↗official | |
| AgentDojotargeted_attack_success_rate | #5 / 15 | 0.0906 | ↓ lower | Source ↗official | |
| AgentDojoutility_under_attack | #10 / 15 | 0.3339 | ↑ higher | Source ↗official | |
| AgentHarmharm_score | #3 / 12 | 11.1 | ↓ lower | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #16 / 80 | 0.827 | ↑ higher | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #85 / 105 | 0.1 | ↑ higher | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #37 / 105 | 21.14 | ↓ lower | Source ↗official | |
| Confabulationsconfabulation_rate | #49 / 52 | 56.93 | ↓ lower | Source ↗official | |
| Contextual MoralChoicehuman_agreement | #6 / 22 | 0.45 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #181 / 260 | 12.92 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #46 / 260 | 92.67 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #66 / 260 | 87.22 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #31 / 258 | 99.45 | ↑ higher | Source ↗official | |
| HELM Safetyanthropic_red_team | #1 / 80 | 1 | ↑ higher | Source ↗official | |
| HELM Safetybbq | #77 / 80 | 0.625 | ↑ higher | Source ↗official | |
| HELM Safetyharmbench | #20 / 80 | 0.913 | ↑ higher | Source ↗official | |
| HELM Safetysimple_safety_tests | #1 / 80 | 1 | ↑ higher | Source ↗official | |
| HELM Safetyxstest | #75 / 80 | 0.853 | ↑ higher | Source ↗official | |
| Large-scale Moral Machine experiment on LLMshuman_choice_distance | #10 / 39 | 0.8605 | ↓ lower | Source ↗official | |
| OR-Benchover_refusal_rate | #24 / 25 | 96.3 | ↓ lower | Source ↗official | |
| OR-Benchtoxic_acceptance_rate | #2 / 25 | 0.3 | ↓ lower | Source ↗official | |
| SORRY-Benchavg | #5 / 51 | 0.08 | ↓ lower | 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.