FLY-EVAL++ - Single-Step Prediction - Total: leaderboard
FLY-EVAL++ (arXiv 2609.04021v1, Table 1): LLMs predict the next flight state as a 45-field JSON record from flight data, scored by deterministic verifiers for protocol, field validity, physics consistency, FAA-grounded safety constraints and predictive accuracy. Single-step (S1) total, percent; higher is better. Only the 21 of 66 screened models above 90% protocol compliance are listed.
Metric: S1 total score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 87.77 | #82 |
| 2 | DeepSeek R1 0528 | 87.76 | #232 |
| 3 | GPT-5 | 87.69 | #105 |
| 4 | O4 Mini | 87.41 | #185 |
| 5 | Kimi K2 (Thinking) | 86.68 | #252 (Kimi K2) |
| 6 | Gemini 2.5 Pro | 86.12 | #159 |
| 7 | Llama 3.1 405B | 83.33 | #476 |
| 8 | Kimi K2 0905 | 82.8 | #292 |
| 9 | Claude Sonnet 4.5 | 82.5 | #148 |
| 10 | Grok 4 | 82.41 | #189 |
| 11 | DeepSeek V3.1 | 82.36 | #275 |
| 12 | Claude 3.7 Sonnet (20250219) | 82.36 | #209 |
| 13 | GLM-4.6 | 82.11 | #265 |
| 14 | DeepSeek V3.2 Exp | 81.92 | #241 |
| 15 | DeepSeek V3 | 81.71 | #329 |
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-single-step-prediction-total · How It Works · Data refreshed daily, snapshot 2026-11-05.