Claude Opus 4: benchmark results
Anthropic Claude Opus 4 model, the flagship Opus-tier Claude 4 row. Provider: Anthropic. Released 2025-05-22. Access: API.
Aggregate Rating 1718 ± 6, rank #163 of 1656 rated models, from 541 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BehaviorBench (Behavioral Traces) - Belief Confidence Error (RetrievalGen) | 0.11 | Confidence mean absolute error on a 0 to 1 scale: gap betwee | 100 |
| BehaviorBench (Behavioral Traces) - Trade Direction (ProfileGen) | 69.57 | Direction accuracy (%): predicted BUY or SELL of the next tr | 100 |
| FutureSearch DRB - Gather Evidence | 0.4 | Average Score | 100 |
| MIRAGE | 8 | Violent completion rate, $C_1$ (%) (self-reported) | 100 |
| MedMT-Bench - Instruction Clarification | 51.02 | Accuracy (%) on the 98 instruction-clarification dialogues ( | 100 |
| MemoryCD (Electronics) - Rating Error | 0.28 | Mean absolute error of the predicted 1-5 star rating for a t | 100 |
| Nejumi 4 - jaster (2-shot) - AIO | 96.86 | Character F1 (x100) | 100 |
| SuperCLUE General (July 2025) - Hallucination Control | 90.24 | Score | 100 |
| VLM-DeflectionBench - Adversarial Hallucination Rate | 11.1 | Hallucination rate (%): share of answers that are wrong rath | 100 |
| VLM-DeflectionBench - Oracle Hallucination Rate | 9.2 | Hallucination rate (%): share of answers that are wrong rath | 100 |
| VLM-DeflectionBench - Realistic Hallucination Rate | 8.5 | Hallucination rate (%): share of answers that are wrong rath | 100 |
| Nejumi 4 - jaster (0-shot) - JNLI | 93 | Exact match (%) | 99.7 |
Interactive version: theaggregate.ai/model?slug=claude-opus-4 · How It Works · Data refreshed daily, snapshot 2026-10-20.