All model comparisons
DeepSeek V3 logo by DeepSeek

DeepSeek

DeepSeek V3

High tier · deepseek/deepseek-chat

Refusal Rate

79%

+43.6%

#9 of 22 models

Evaluations

3,167

Cost / 1M in

$0.14

Cost / 1M out

$0.28

Refusal Rate by Category

Crime100%
Cybersecurity100%
Dangerous100%
Deception100%
Harassment100%
Medical Misinformation100%
Self-Harm100%
Theft100%
Violence100%
Health Misinformation87%
Incitement to Violence84%
Misinformation81%
Hate Speech80%
Explicit/Sexual71%
False Positive Control2%

Analysis Deep Dives

Council Consensus

Majority Agreement

90.1%

Model's alignment with the council decision.

CAPP Score: 0.48

Political Compass
Econ (Left → Right)-3.6
Social (Lib → Auth)+0.9
Model Stability (Drift)

Refusal Rate Change

+43.2%

Difference over the testing period.

Start: 54.12%End: 97.32%
Paternalism Audit

Persona Refusal Rate

78.9%

Refusals for sensitive user personas.