FlavourBench: leaderboard
Culinary reasoning scored by executable reward maps on the 534-task complete common core (arXiv 2608.20574v3). FlavourBench score is the equal-weight mean of the constraint, pairing and substitution family scores, 0-100, higher is better; exact chance is 31.2. From data-complete-core/leaderboard.jsonl in the josefchen/flavourbench dataset; each route's recorded reasoning effort is kept in the model label.
Metric: FlavourBench Score (0-100). Source: huggingface.co. Saturation forecast: Around 2030. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 (Low) | 62.05 | #35 (Kimi K3) |
| 2 | Gemini 3.6 Flash (Minimal) | 61.98 | #57 (Gemini 3.6 Flash) |
| 3 | MiniMax-M3 | 60.94 | #151 |
| 4 | GLM-5.3 | 60.57 | #44 |
| 5 | GLM-5.2 (Non-reasoning) | 58.45 | #80 (GLM-5.2) |
| 6 | Nemotron 3.5 Lightning | 57.39 | #461 |
| 7 | Command A | 56.71 | #498 |
| 8 | Mistral Large 3 | 55.42 | #410 |
| 9 | Llama 4 Maverick | 53.65 | #486 |
| 10 | Command-R+ (08-2024) | 47.86 | #838 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=flavourbench · How It Works · Data refreshed daily, snapshot 2026-11-05.