CausalGame - Agent Mode: leaderboard
Causal discovery by experiment in 14 scenarios modelled on historical cases: the model designs experiments on a hidden causal system with selection bias, confounders and noisy measurements, then commits to a decision, scored by the survival rate it achieves. Agent mode: structured tool calls in a mandatory ReAct loop, with sandboxed Python for analysis; mean of three seeds. Higher is better. Source: arxiv.org/abs/2607.04293.
Metric: Mean Survival Rate (%; 14 scenarios, 3 seeds). Source: arxiv.org. Saturation forecast: Around 2030. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 68 | #90 |
| 2 | GLM-5.2 | 67.8 | #82 |
| 3 | GPT-5.5 (High) | 67.8 | #34 (GPT-5.5) |
| 4 | Qwen 3.7 Max | 66.9 | #88 |
| 5 | GPT-5.5 | 66.8 | #34 |
| 6 | GLM-5.1 | 66.5 | #111 |
| 7 | Claude Opus 4.7 | 65.4 | #53 |
| 8 | Grok 4.1 Fast | 65.4 | #227 |
| 9 | GPT-5.5 (xHigh) | 65.4 | #34 (GPT-5.5) |
| 10 | DeepSeek V4 Pro | 64.6 | #113 |
| 11 | Kimi K2.6 | 64.6 | #115 |
| 12 | MiniMax-M2 | 64.5 | #329 |
| 13 | Claude Sonnet 4.5 | 64.4 | #150 |
| 14 | Gemini 3.5 Flash | 64.1 | #68 |
| 15 | MiMo-V2.5-Pro | 63.4 | #169 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=causalgame-agent-mode · How It Works · Data refreshed daily, snapshot 2026-11-07.