Claude 3.7 Sonnet: benchmark results
Anthropic's first hybrid-reasoning Sonnet, switching between near-instant replies and extended thinking with an API thinking budget (February 2025). Provider: Anthropic. Released 2025-02-24. Access: API.
Aggregate Rating 1679 ± 4, rank #249 of 1656 rated models, from 1486 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AGC-Bench - riddlesense | 1.25 | Dataset z-score | 100 |
| AgentRewardBench (Accessibility Tree) - AssistantBench | 87.5 | Precision (%) | 100 |
| AgentRewardBench (Accessibility Tree) - WorkArena++ | 66.7 | Precision (%) | 100 |
| AgentRewardBench (Screenshot) | 69.4 | Precision (%) | 100 |
| AgentRewardBench (Screenshot) - VisualWebArena | 64.8 | Precision (%) | 100 |
| AgentRewardBench (Screenshot) - WorkArena++ | 66.7 | Precision (%) | 100 |
| BehaviorBench (Behavioral Traces) - Belief Confidence Error (ProfileGen) | 0.08 | Confidence mean absolute error on a 0 to 1 scale: gap betwee | 100 |
| CodeGuard | 64 | F1 (%; three-way classification of 1,000 manually verified t | 100 |
| Complexity Ceiling Benchmark | 31.9 | Accuracy (%; share of parseable trials whose final answer an | 100 |
| Complexity Ceiling Benchmark - Symbolic Pointer Tracking | 71.2 | Accuracy (%; share of parseable trials whose final answer an | 100 |
| DARE-bench (Data Science) - Classification (ML Modeling) | 61.03 | Macro-F1 (shown times 100) against the ground-truth labels o | 100 |
| DARE-bench (Data Science) - Regression (ML Modeling) | 63.2 | Clipped R-squared (0 to 1, shown times 100) against the grou | 100 |
Interactive version: theaggregate.ai/model?slug=claude-3-7-sonnet · How It Works · Data refreshed daily, snapshot 2026-10-20.