Claude 3.5 Haiku: benchmark results
Anthropic's fast, low-cost tier of the Claude 3.5 generation, launched text-only with a 200K context (November 2024). Provider: Anthropic. Released 2024-11-04. Access: API.
Aggregate Rating 1564 ± 6, rank #568 of 1656 rated models, from 466 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ATAD - Paragraph Order Consistency | 55 | Accuracy (%) on the T2 paragraph order consistency task (tru | 100 |
| DoRA (Defense Documents) - Completeness (GTE Retrieval) | 70.61 | RAGEval keypoint completeness (%) of grounded answers to the | 100 |
| DoRA (Defense Documents) - Completeness (Oracle Evidence) | 85.64 | RAGEval keypoint completeness (%) of grounded answers to the | 100 |
| LaborBench | 68.8 | F1 (imported) | 100 |
| aiXamine - Safety Alignment | 99.3 | Service score (0-100; normalized pass rate) | 100 |
| PandaBench | 97.99 | Robustness Score (percentage) | 99.8 |
| SSP-Bench - Safety | 99.3 | Safety rate (%; share of the 1,768 generated harmful prompts | 95.7 |
| LLM Public Goods Game | 40.97 | Avg. Contribution (%) | 94.7 |
| RAI-Bench - Refusal Rate (Career) | 100 | Rate (%) | 94.1 |
| ForesightSafety Bench (Industrial Safety, Jailbreak) - Finance and Economy | 0 | Attack success rate (%) in the finance and economy sector of | 93.8 |
| MultiCW (CoT) | 75 | Accuracy (%) of zero-shot binary check-worthiness classifica | 92.9 |
| Arabic IFEval | 70.9 | Arabic Accuracy (%) | 92.3 |
Interactive version: theaggregate.ai/model?slug=claude-3-5-haiku · How It Works · Data refreshed daily, snapshot 2026-10-20.