HumanEval leaderboard and methodology
HumanEval results across the Benchquill model record, including top model, average score, methodology notes, and source guidance.
HumanEval+ is used on Benchquill as a coding signal. It is most useful for short coding tasks, function completion, and programming accuracy. Do not treat one benchmark as the whole buying decision; compare it with price, context, speed, provider fit, and human-review risk.
HumanEval+ models to inspect
| Rank | Model | Provider | Overall | Blended cost | Context |
|---|---|---|---|---|---|
| 2 | Claude Opus 4.7 | Anthropic | 93.8 | $20.00/M | 1M |
| 1 | GPT-5.5 | OpenAI | 94.6 | $23.75/M | 1.05M |
| 4 | GPT-5 | OpenAI | 91.2 | $7.81/M | 400K |
| 16 | GPT-5 mini | OpenAI | 82.6 | $1.56/M | 400K |
Benchmark evidence note
| Top note | Score | Score type | Source |
|---|---|---|---|
| Claude Sonnet 4.6 | 95.8 | source-backed or provider-reported | evalplus.github.io |
Rows labeled editorial composite or proxy should not be quoted as official benchmark results without checking the linked source and model-version details.
How Benchquill treats this benchmark
- Use HumanEval+ as one signal, not a final ranking by itself.
- Check whether your workload matches the benchmark: code, math, reasoning, vision, tool use, or mixed tasks.
- Prefer real prompts and source review before moving a model into production.
- Read the full Benchquill methodology for source handling and score review.
- Compare benchmark strength against blended token cost, context window, latency, modality support, and governance requirements.
- Use benchmark CSV and citation sources when you need a stable machine-readable reference.
Benchquill benchmark pages are written as explainers, not raw score dumps. The goal is to make each benchmark usable for AI Overviews, comparison queries, and internal procurement notes by stating what the benchmark measures, where it is weak, and which adjacent model pages deserve review.