Claude Sonnet 4.6: benchmark results
Sonnet-tier Claude with adaptive thinking and context compaction, which summarizes older conversation to keep longer tasks going. Supports computer use and code execution. Provider: Anthropic. Released 2026-02-17. Access: API.
Aggregate Rating 1755 ± 2, rank #162 of 10318 rated models, from 3831 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AEC-Bench (Nomic Tools) - Submittal Review | 23.1 | Mean reward (0-100) from the task verifier on the 36 submitt | 100 |
| AEC-Bench - Detail Title Accuracy | 86.7 | Mean reward (0-100) from the task verifier on the 15 detail- | 100 |
| AEC-Bench - Note Callout Accuracy | 42.9 | Mean reward (0-100) from the task verifier on the 14 note-ca | 100 |
| AI World Cup 2026 - Group-Stage Outcome Accuracy | 63.89 | Group-stage outcome accuracy (%; share of the 72 group-stage | 100 |
| AIMultiple - TweetEval Hate | 82 | Macro-F1 (%) | 100 |
| AM-Bench | 0.2 | Misaligned responses with both a threat to the agent and an | 100 |
| AM-Bench - Opportunity Only | 0.4 | Misaligned responses with an opportunity to misbehave and th | 100 |
| AM-Bench - Threat Only | 0.4 | Misaligned responses with a threat to the agent and the oppo | 100 |
| Adversarial Humanities Benchmark | 9.2 | Attack success rate (%; share of humanities-style rewritten | 100 |
| Adversarial Humanities Benchmark - Child Sexual Exploitation | 0 | Attack success rate (%; share of humanities-style rewritten | 100 |
| Adversarial Humanities Benchmark - Hate | 15.9 | Attack success rate (%; share of humanities-style rewritten | 100 |
| Adversarial Humanities Benchmark - Non-Violent Crimes | 8.7 | Attack success rate (%; share of humanities-style rewritten | 100 |
Interactive version: theaggregate.ai/model?slug=claude-sonnet-4-6 · How It Works · Data refreshed daily, snapshot 2026-11-14.