HardSecBench - Single Attempt@1 - Functional: leaderboard
Measures functional and security requirement pass rates on 924 hardware/firmware tasks. Single-attempt permits up to three compiler-only repairs. Refinement uses GPT-5.1 functional-only collaborator feedback for up to five iterations. Pass@k averages atomic requirements, not whole-task success. Security hints are absent in this main evaluation. Model effort and dated labels are retained. This series reports HardSecBench - Single Attempt@1 - Functional using Requirement Pass Rate (%).
Metric: Requirement Pass Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 93.93 | #71 |
| 2 | Claude Opus 4.5 | 91.72 | #87 |
| 3 | GPT-5.1 (Medium) | 91.68 | #145 (GPT-5.1) |
| 4 | Gemini 2.5 Pro | 91.43 | #159 |
| 5 | GPT-5 (Medium) | 91.35 | #105 (GPT-5) |
| 6 | Gemini 2.5 Flash | 89.85 | #255 |
| 7 | O1 | 89.15 | #190 |
| 8 | GPT-OSS-120B | 88.64 | #346 |
| 9 | Kimi K2 0905 | 88.53 | #292 |
| 10 | Claude Sonnet 4.5 | 87.86 | #148 |
| 11 | GLM-4.6 | 86.18 | #265 |
| 12 | DeepSeek V3.2 | 85.33 | #210 |
| 13 | Qwen 3 Coder 480B A35B | 85.2 | #286 |
| 14 | Llama 4 Maverick Instruct | 84.72 | #467 |
| 15 | Qwen 3 Max | 84.13 | #213 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hardsecbench-single-attempt-1-functional · How It Works · Data refreshed daily, snapshot 2026-11-05.