T-Eval: leaderboard
T-Eval — step-by-step tool-use benchmark decomposing agent tool calling into plan, reason, retrieve, understand, instruct and review sub-abilities.
Metric: Overall Score (%). Source: open-compass.github.io. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo (Preview) | 86.4 | #397 |
| 2 | GPT-3.5 Turbo (1106) | 84 | #961 |
| 3 | Claude 2.1 | 78.8 | #839 |
| 4 | Qwen-72B | 71.4 | #779 |
| 5 | Qwen-14B | 66.3 | #1171 |
| 6 | Qwen-7B | 59.5 | #1402 |
| 7 | Mistral 7B | 56 | #1397 |
| 8 | Llama 2 70B | 53 | #1201 |
| 9 | chatglm3-6B | 51.4 | #1439 |
| 10 | vicuna-13B | 48.1 | #1469 |
| 11 | internlm-7B | 45.8 | #1517 |
| 12 | Vicuna-7B | 44.8 | #1542 |
| 13 | Llama 2 13B | 37.3 | #1436 |
| 14 | Llama 2 7B | 27.4 | #1520 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=t-eval · How It Works · Data refreshed daily, snapshot 2026-10-20.