HardEval - All: leaderboard
HardEval (2026), temperature-zero Bash hardening scripts scored by a model-agnostic grader on AlmaLinux 8, OpenSCAP 1.3.14, SCAP Security Guide 0.1.81 STIG profile; script extraction accepts fenced or raw Bash; fixed test denominator 151; printed FP16 and 4-bit configurations remain separate. Execution-verified Compliance (%); higher is better.
Metric: Execution-verified Compliance (%). Source: arxiv.org. Saturation forecast: Around January 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 80.8 | #49 |
| 2 | Claude Opus 4.8 | 76.8 | #37 |
| 3 | Claude Sonnet 4.5 | 75.5 | #142 |
| 4 | Claude Opus 4.5 | 72.8 | #83 |
| 5 | GPT-4o | 66.2 | #343 |
| 6 | Claude Sonnet 5 | 58.9 | #57 |
| 7 | Claude Haiku 4.5 | 50.3 | #280 |
| 8 | Qwen 2.5 Coder 14B Instruct | 35.1 | #675 |
| 9 | Qwen 2.5 Coder 7B Instruct | 15.9 | #951 |
| 10 | glm-4-9B | 13.9 | #889 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hardeval-all · How It Works · Data refreshed daily, snapshot 2026-10-20.