HumanEval
Code generation benchmark. 164 hand-written Python problems with unit tests.
Category: coding ·
Metric: pass@1 ·
Source: github.com ↗
Leaderboard
| Rank | Model | Provider | Score | Measured | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | Anthropic | 92.0 | 2026-03-12 | ↗ |
| 2 | Mistral Large 2 | Mistral AI | 92.0 | 2024-07-24 | ↗ |
| 3 | GPT-4o | OpenAI | 90.2 | 2024-05-13 | ↗ |
| 4 | Llama 3.3 70B | Meta | 88.4 | 2024-12-06 | ↗ |
| 5 | DeepSeek-V3 | DeepSeek | 82.6 | 2024-12-26 | ↗ |
What this benchmark measures, in detail
Code generation benchmark. 164 hand-written Python problems with unit tests.
Different benchmarks measure different things. A model that excels on HumanEval may underperform on real-world workloads if the benchmark's distribution doesn't match your data. Use benchmark scores as a triage signal — narrow to a shortlist — then evaluate on your actual workload before committing.
Methodology notes
Scores in the leaderboard are taken from the model's release announcement or model card, cited via the "Source" link. Where two sources disagree (which happens often for SWE-bench and IFEval), the linked primary source wins. Reproducibility for some benchmarks (notably anything graded by an LLM) varies by run — treat the score as ±2-3 points unless the source is a peer-reviewed result.