CRYSTAL: leaderboard
Metric: Match F1 (0-1, shown times 100): F1 of the model's reasoning steps against the reference steps, greedy one-to-one matching at cosine similarity 0.35 with all-distilroberta-v1, on the 6,372-question CRYSTAL test set, greedy decoding, JSON output of reasoning steps and answer (malformed outputs keep placeholder steps); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 77.3 | #176 |
| 2 | Gemini 2.5 Flash | 67.3 | #237 |
| 3 | Qwen 2.5 VL 32B Instruct | 65.3 | #443 |
| 4 | Gemma 3 4B | 61.8 | #1084 |
| 5 | GPT-5 | 61.2 | #91 |
| 6 | Gemma 3 12B | 60.5 | #666 |
| 7 | GPT-5.2 Instant | 56.4 | #205 |
| 8 | Qwen 2.5 VL 7B Instruct | 47.5 | #643 |
| 9 | Llama 3.2 11B | 47.1 | #1183 |
| 10 | MiniCPM-V-2.6 | 21.5 | #825 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=crystal · How It Works · Data refreshed daily, snapshot 2026-10-11.