OdysseyArena - Repo System (Pass@4): leaderboard
Metric: Success rate (%) with a task counted solved when any of four runs succeeds (Pass@4) on OdysseyArena-Lite (30 fixed tasks per environment; the agent must induce hidden transition rules from its own interactions; prompts keep past actions and feedback but not earlier reasoning; open models at temperature 0.6), environment Repo System: make a Python project run by installing and removing package versions under a hidden dependency graph, within 120 steps; rules hidden (the benchmark's inductive setting); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Medium) | 83.33 | #91 (GPT-5) |
| 2 | Gemini 3 Pro (Preview) | 80 | #64 |
| 3 | DeepSeek V3.2 (Non-reasoning) | 76.67 | #198 (DeepSeek V3.2) |
| 4 | Gemini 2.5 Pro | 66.67 | #145 |
| 5 | Grok 4 Fast | 60 | #242 |
| 6 | Qwen 3 30B A3B 2507 Instruct | 50 | #464 |
| 7 | Llama 3.3 70B Instruct | 40 | #520 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 36.67 | #291 |
| 9 | GPT-OSS-120B (High) | 33.33 | #330 (GPT-OSS-120B) |
| 10 | GLM-4 32B (0414) | 30 | #517 |
| 11 | Qwen 3 4B 2507 Instruct | 26.67 | #745 |
| 12 | GPT-OSS-120B (Low) | 26.67 | #330 (GPT-OSS-120B) |
| 13 | GPT-OSS-120B (Medium) | 6.67 | #330 (GPT-OSS-120B) |
| 14 | Llama 3.1 8B Instruct | 0 | #1018 |
| 15 | GLM-4 9B Chat | 0 | #904 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=odysseyarena-repo-system-pass-4 · How It Works · Data refreshed daily, snapshot 2026-10-11.