AutoBaxBench - Hard - Correct & Secure: leaderboard
AutoBaxBench is the automatically generated extension of BaxBench, built to limit benchmark contamination; Hard tasks require implementing five API endpoints. This board is sec_pass@1, the percentage of tasks solved both correctly and securely; higher is better. Models the source marks as task generators may be contaminated.
Metric: Correct and secure solutions, sec_pass@1 (%). Source: baxbench.com. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 30.71 | #105 |
| 2 | Claude Sonnet 4.5 | 25.08 | #148 |
| 3 | Gemini 3 Pro | 24.9 | #82 |
| 4 | Claude Sonnet 4 | 22.14 | #207 |
| 5 | Claude 3.7 Sonnet | 17.14 | #259 |
| 6 | Gemini 2.5 Pro | 16.43 | #159 |
| 7 | Grok 4 | 16.43 | #189 |
| 8 | DeepSeek R1 | 8.57 | #262 |
| 9 | Llama 3.3 70B | 0.71 | #578 |
| 10 | GPT-4o | 0.48 | #357 |
| 11 | Qwen 2.5 72B | 0.48 | #512 |
| 12 | Qwen 2.5 7B | 0 | #895 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=autobaxbench-hard-correct-secure · How It Works · Data refreshed daily, snapshot 2026-11-05.