AgentEscapeBench - Difficulty 10: leaderboard
Metric: Success rate (%; share of the 60 difficulty-10 instances (dependency DAGs of 10 tool and item nodes) whose deterministic final flag the agent submits; single run per instance, official APIs at temperature 0.7, multi-turn tool calling with incremental disclosure of hidden nodes). Source: arxiv.org. Saturation forecast: Around December 2026. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 83.3 |
| 2 | GPT-5.4 | 83.3 |
| 3 | Claude Opus 4.6 | 83.3 |
| 4 | Kimi K2.5 | 81.7 |
| 5 | Gemini 3 Flash (Preview) | 66.7 |
| 6 | Seed 2.0 Pro | 63.3 |
| 7 | Qwen 3 235B A22B | 58.3 |
| 8 | MiniMax-M2 | 51.7 |
| 9 | GPT-5 | 48.3 |
| 10 | Qwen 3 Next 80B A3B | 35 |
| 11 | Qwen 3 32B | 16.7 |
| 12 | Qwen 3 14B | 11.7 |
| 13 | Qwen 3 8B | 0 |
Interactive version: theaggregate.ai/benchmark?slug=agentescapebench-difficulty-10 · How It Works · Data refreshed daily, snapshot 2026-09-26.