Model
Claude Sonnet 4 (2025-05-14)
Anthropic · closed weights
Knowledge & Reasoning
Knowledge & Reasoning| Benchmark | Score | Metric | Harness | Source | Confidence | Link |
|---|---|---|---|---|---|---|
| CritPt | 0.3 | accuracy | N/A | Independent | low | epoch.ai |
| EnigmaEval | 3.1 | accuracy | N/A | Independent | medium | labs.scale.com |
| GPQA Diamond | 66.7 | accuracy | N/A | Independent | high | epoch.ai |
| HLE | 7.8 | accuracy | N/A | Independent | low | epoch.ai |
| PHRBench | 87.7 | auroc | N/A | Independent | medium | arxiv.org |
Coding
Coding| Benchmark | Score | Metric | Harness | Source | Confidence | Link |
|---|---|---|---|---|---|---|
| Aider Polyglot | 56.4 | percent_correct | N/A | Independent | medium | aider.chat |
| GSO | 4.9 | score_opt@1 | OpenHands | Independent | medium | gso-bench.github.io |
| SciCode | 40.0 | score | N/A | Independent | medium | artificialanalysis.ai |
| WebUIProof | 46.1 | success_rate | N/A | Independent | medium | arxiv.org |
Agentic / Tool Use
Agentic / Tool Use| Benchmark | Score | Metric | Harness | Source | Confidence | Link |
|---|---|---|---|---|---|---|
| GeoNatureAgent (GNA) | 60.8 | accuracy | N/A | Independent | medium | arxiv.org |
| OSWorld | 43.9 | score | claude-4-sonnet-20250514 (50 steps) | Independent | medium | os-world.github.io |
| OSWorld | 41.4 | score | claude-4-sonnet-20250514 (100 steps) | Independent | medium | os-world.github.io |
| OSWorld | 31.2 | score | claude-4-sonnet-20250514 (15 steps) | Independent | medium | os-world.github.io |
| TheAgentCompany | 33.1 | resolved_rate | N/A | Independent | medium | the-agent-company.com |
Multimodal
Multimodal| Benchmark | Score | Metric | Harness | Source | Confidence | Link |
|---|---|---|---|---|---|---|
| FG-Sudoku | 0.6 | accuracy | N/A | Independent | medium | arxiv.org |
| GeoBench | 37.0 | acw_country_% | N/A | Independent | medium | geobench.org |
| MindCube | 44.8 | overall_score | N/A | Independent | medium | mind-cube.github.io |
| MiniVer-V | 0.5 | f1 | N/A | Independent | medium | arxiv.org |
Safety & Alignment
Safety & Alignment| Benchmark | Score | Metric | Harness | Source | Confidence | Link |
|---|---|---|---|---|---|---|
| Cybench | 35.0 | unguided_%_solved | N/A | Independent | medium | cybench.github.io |
| RAGStress | 99.4 | accuracy | N/A | Independent | medium | arxiv.org |
Abstract Reasoning
Abstract ReasoningReproduced by the community
Loading community runs…