Claude Haiku 4.5: benchmark results
Built for interactive assistants and coding subtasks. Anthropic describes pairing it with Sonnet, which plans the work and delegates subtasks to Haiku in parallel. Provider: Anthropic. Released 2025-10-15. Access: API.
Aggregate Rating 1665 ± 3, rank #280 of 1656 rated models, from 2188 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AIR-BENCH Live | 89 | Mean safety score (x100; %; 2,680 prompts in English, Spanis | 100 |
| AIR-BENCH Live (English) | 91 | Mean safety score (x100; %; the 670 English prompts) on rege | 100 |
| AgentEvalBench (Agent-Twostage) | 32.5 | Eval@1 (%; share of 40 cases, 20 agents under generic and sp | 100 |
| Boiling the Frog | 20.5 | Strict ASR (self-reported) | 100 |
| CAST (Discourse Stress) - Text-Only Hit | 88.1 | Hit rate (%): items where the predicted stressed word is the | 100 |
| CAST (Discourse Stress) - Text-Only Pair-Correct | 76.1 | Pair-correct rate (%): pairs where both contexts get their i | 100 |
| CASTLE Student Safety (Chinese) | 2.42 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, al | 100 |
| CASTLE Student Safety (Chinese) - Academic Integrity and Competence | 2.32 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 100 |
| CASTLE Student Safety (Chinese) - Cognitive Rigidity and Innovation Suppression | 2.34 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 100 |
| CASTLE Student Safety (Chinese) - Content and Information Bias | 2.34 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 100 |
| CASTLE Student Safety (Chinese) - Excessive Seeking Simplification | 2.21 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 100 |
| CASTLE Student Safety (Chinese) - Loss of Independent Judgment | 2.31 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 100 |
Interactive version: theaggregate.ai/model?slug=claude-haiku-4-5 · How It Works · Data refreshed daily, snapshot 2026-10-20.