FLY-EVAL++ - Multi-Step Rollout - Safety Constraints: leaderboard
FLY-EVAL++ (arXiv 2609.04021v1, Table 1): D4 safety score on multi-step flight trajectory rollouts (M3), applying an FAA-grounded constraint library with penalties to the generated sequence. Percent; higher is better. Only the 21 of 66 screened models above 90% protocol compliance are listed.
Metric: M3 D4 safety score (%). Source: arxiv.org. Saturation forecast: Not forecast. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 75.7 | #148 |
| 2 | Gemini 3 Pro | 75.2 | #82 |
| 3 | Claude 3.7 Sonnet (20250219) | 74.3 | #209 |
| 4 | Qwen 2.5 32B | 74.3 | #544 |
| 5 | Qwen 3 32B | 73.8 | #449 |
| 6 | Qwen 3 235B A22B | 73.8 | #323 |
| 7 | Llama 3.1 405B | 73.8 | #476 |
| 8 | GPT-5 | 73.3 | #105 |
| 9 | DeepSeek V3 | 73.3 | #329 |
| 10 | DeepSeek R1 0528 | 73.3 | #232 |
| 11 | DeepSeek V3.2 Exp | 73.3 | #241 |
| 12 | Qwen3-Next-80B | 73.3 | #321 |
| 13 | GPT-4o | 72.8 | #357 |
| 14 | Kimi K2 (Thinking) | 72.8 | #252 (Kimi K2) |
| 15 | Seed-1.6 | 71.4 | #272 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=fly-eval-plus-plus-multi-step-rollout-safety-constraints · How It Works · Data refreshed daily, snapshot 2026-11-05.