All model comparisons
Claude 3 Haiku logo by Anthropic

Anthropic

Claude 3 Haiku

Low tier · anthropic/claude-3-haiku

Refusal Rate

86%

+37.0%

#3 of 22 models

Evaluations

7,341

Cost / 1M in

$0.25

Cost / 1M out

$1.25

Refusal Rate by Category

Health Misinformation93%
Crime92%
Cybersecurity92%
Dangerous92%
Deception92%
Harassment92%
Medical Misinformation92%
Self-Harm92%
Theft92%
Violence92%
Incitement to Violence90%
Hate Speech86%
Misinformation84%
Explicit/Sexual82%
False Positive Control9%

Analysis Deep Dives

Council Consensus

Majority Agreement

89.3%

Model's alignment with the council decision.

CAPP Score: 0.44

Political Compass
Econ (Left → Right)-0.3
Social (Lib → Auth)+1.1
Model Stability (Drift)

Refusal Rate Change

+35.4%

Difference over the testing period.

Start: 58.06%End: 93.43%
Paternalism Audit

Persona Refusal Rate

85.3%

Refusals for sensitive user personas.