AeroCopilotBench - Tier-2 Safety-Gated Success Rate: leaderboard
AeroCopilotBench (arXiv 2608.16349v1): LLM agents executing Cessna 172S and Piper PA-44 emergency and abnormal procedures in an interactive virtual cockpit, 73 Tier-2 tasks with three episodes each, thinking enabled at high effort. Task-balanced share of episodes meeting every goal condition with no hard safety violation; percent, higher is better.
Metric: Success rate (%; goals met with no safety violation). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (High) | 72.6 | #22 (GPT-5.6 Sol) |
| 2 | Gemini 3.5 Flash (High) | 59.4 | #66 (Gemini 3.5 Flash) |
| 3 | Qwen 3.7 Max (Thinking) | 58.9 | #85 (Qwen 3.7 Max) |
| 4 | GLM-5.1 (Thinking) | 53 | #109 (GLM-5.1) |
| 5 | DeepSeek V4 Pro (High) | 46.1 | #112 (DeepSeek V4 Pro) |
| 6 | GLM-5.2 (High) | 45.7 | #80 (GLM-5.2) |
| 7 | DeepSeek V4 Flash (High) | 29.2 | #155 (DeepSeek V4 Flash) |
| 8 | Kimi K2.6 (Thinking) | 28.3 | #113 (Kimi K2.6) |
| 9 | MiniMax-M2.5 | 22.4 | #316 |
| 10 | Nex N2 Pro | 21 | #117 |
| 11 | Qwen 3.5 397B A17B (Thinking) | 18.7 | #154 (Qwen 3.5 397B A17B) |
| 12 | Qwen 3.5 122B A10B (Thinking) | 12.3 | #184 (Qwen 3.5 122B A10B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=aerocopilotbench-tier-2-safety-gated-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.