Axion 1.0 - BioMysteryBench Human-Difficult: leaderboard
Axion 1.0 author-run ReAct environment with web search, webpage visits, scholarly retrieval and Python. Best single run of three, or five for BioMysteryBench human-difficult; benchmark-approved website restrictions apply. Full HLE restricted websites and HLE-130 text-only are separate populations. Externally reported baseline scores are not pooled. BioMysteryBench Human-Difficult.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Axion Run: harness=ReAct; selection=best%20of%205) | 47.1 | |
| 2 | Axion 1.0-122B (Axion Run: harness=ReAct; selection=best%20of%205) | 41.7 | |
| 3 | Gemini 3.1 Pro (Axion Run: harness=ReAct; selection=best%20of%205) | 35.3 | |
| 4 | Axion 1.0-35B (Axion Run: harness=ReAct; selection=best%20of%205) | 35.3 | |
| 5 | Axion 1.0-9B (Axion Run: harness=ReAct; selection=best%20of%205) | 23.5 | |
| 6 | Qwen3.6-35B-A3B (Axion Run: harness=ReAct; selection=best%20of%205) | 23.5 | |
| 7 | Qwen3.5-122B-A10B (Axion Run: harness=ReAct; selection=best%20of%205) | 17.6 | |
| 8 | Axion 1.0-4B (Axion Run: harness=ReAct; selection=best%20of%205) | 11.8 |
Interactive version: theaggregate.ai/benchmark?slug=axion-1-0-biomysterybench-human-difficult · How It Works · Data refreshed daily, snapshot 2026-11-05.