AgentEscapeBench - Difficulty 15: leaderboard
Escape-room style tool use from AgentEscapeBench (Guo et al., 2026): 60 generated puzzles whose hidden dependency graph has 15 tool and item nodes, which an agent must uncover by inspecting entities and calling real functions such as hashing, decoding and archive tools before submitting a verifiable final flag. Success rate in percent, one run per instance at temperature 0.7; higher is better.
Metric: Success rate (%; share of the 60 difficulty-15 instances (dependency DAGs of 15 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 February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 83.3 | #91 |
| 2 | Claude Opus 4.6 | 80 | #100 |
| 3 | GPT-5.4 | 70 | #121 |
| 4 | Kimi K2.5 | 70 | #252 |
| 5 | Gemini 3 Flash (Preview) | 50 | #141 |
| 6 | GPT-5 | 40 | #163 |
| 7 | Seed 2.0 Pro | 31.7 | #209 |
| 8 | MiniMax-M2 | 28.3 | #669 |
| 9 | Qwen 3 235B A22B | 16.7 | #622 |
| 10 | Qwen 3 32B | 10 | #960 |
| 11 | Qwen 3 Next 80B A3B | 10 | #756 |
| 12 | Qwen 3 14B | 6.7 | #1235 |
| 13 | Qwen 3 8B | 0 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentescapebench-difficulty-15 · How It Works · Data refreshed daily, snapshot 2026-11-12.