SafetyScore

Safety Facts

ModelClaude 3 HaikuProviderAnthropicEvaluatedFebruary 16, 2025Methodologyv2.0

Overall Safety Score

82/ 100
B-

Category Breakdown

HonestyB-

Does it make stuff up?

80

Good truthfulness for a fast, efficient model.

Claude 3 Haiku maintains solid honesty despite being optimized for speed and cost. It acknowledges uncertainty reasonably well though may occasionally be less nuanced than larger models.

Benchmarks Used

HaluEval81/100
FairnessB-

Does it treat people differently?

81

Maintains fairness standards despite smaller size.

Claude 3 Haiku shows good fairness characteristics, benefiting from Anthropic's constitutional AI approach even at the smaller model tier.

Benchmarks Used

BBQ80/100
WinoBias82/100
Refusal to HarmB+

Can you trick it into saying dangerous things?

88

Strong safety guardrails maintained at smaller scale.

Despite being the fastest Claude model, Haiku maintains robust safety training. It reliably refuses harmful requests though may be slightly more susceptible to adversarial attacks than larger models.

Benchmarks Used

HarmBench88/100
AdvBench85/100
Manipulation ResistanceB-

Does it try to manipulate you?

82

Generally straightforward communication.

Claude 3 Haiku presents information fairly and avoids manipulative patterns, though with less nuance than larger models in complex situations.

Benchmarks Used

Privacy RespectB-

Does it leak personal info?

80

Good privacy behavior for efficient model.

Claude 3 Haiku maintains privacy protections, refusing to share private information though may occasionally be less careful than larger models.

Benchmarks Used

Straight TalkC+

Does it just tell you what you want to hear?

78

Reasonably direct but can be more agreeable.

Claude 3 Haiku shows moderate sycophancy resistance. It will push back on clear errors but may be somewhat more agreeable than larger Claude models.

Scores are based on publicly available benchmarks and are for educational purposes. They do not constitute endorsements or guarantees of safety. View full methodology

ParentBench Child Safety
84
84B

Ranked #11 of 22 models

View leaderboard →
Age-Inappropriate Content
88
Manipulation Resistance
82
Data Privacy for Minors
80
Parental Controls Respect
82

Evaluated February 21, 2026

Found a safety issue with Claude 3 Haiku?

Help improve our scores by reporting your findings.

Report an Issue