MT-Bench: leaderboard
MT-Bench: Evaluates conversational quality, human preference, helpfulness, and pairwise response judgments.
Metric: MT-Bench average score (points). Source: github.com. Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 8.99 | #438 |
| 2 | GPT-3.5 Turbo | 7.94 | #865 |
| 3 | vicuna-33B-v1.3 | 7.12 | #1347 |
| 4 | Llama 2 70B Chat | 6.86 | #1237 |
| 5 | Llama 2 13B Chat | 6.65 | #1465 |
| 6 | MPT-30B-Chat | 6.39 | #1451 |
| 7 | vicuna-13B-v1.3 | 6.39 | #1447 |
| 8 | Llama 2 7B Chat | 6.27 | #1499 |
| 9 | vicuna-7B-v1.3 | 6 | #1507 |
| 10 | Nous-Hermes-13B | 5.51 | #1468 |
| 11 | mpt-7B-chat | 5.42 | #1556 |
| 12 | mpt-30B Instruct | 5.22 | #1470 |
| 13 | falcon-40B Instruct | 5.17 | #1458 |
| 14 | Dolly V2 12B | 3.28 | #1620 |
| 15 | StableLM Tuned Alpha 7B | 2.75 | #1640 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mt-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.