ToolBench Leaderboard: leaderboard
ToolBench action-generation leaderboard: the primary eight-task average covers weather, search, booking, spreadsheets, VirtualHome, WebShop Short and tabletop tasks. WebShop Long is reported separately and omitted from the average; ToolBench-tuned checkpoints remain distinct.
Metric: Average Task Score. Source: huggingface.co. Saturation forecast: Estimated already saturated. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 68.76 | #438 |
| 2 | text-davinci-003 | 67.24 | #1146 |
| 3 | CodeLlama-34B-Instruct-hf | 64.79 | #1381 |
| 4 | CodeLlama-34B-hf | 62.94 | #1515 |
| 5 | Llama 2 70B | 61.03 | #1201 |
| 6 | CodeLlama-13B-Instruct-hf | 60.49 | #1532 |
| 7 | CodeLlama-34B-Python-hf | 59.16 | #1552 |
| 8 | GPT-3.5 Turbo | 56.65 | #865 |
| 9 | CodeLlama-13B-Python-hf | 56.31 | #1623 |
| 10 | LLaMA-65B | 55.59 | #1303 |
| 11 | CodeLlama-7B-Python-hf | 52.15 | #1608 |
| 12 | CodeLlama-7B-Instruct-hf | 50.48 | #1545 |
| 13 | starcoder | 49.75 | #1577 |
| 14 | LLaMA-30B | 49.59 | #1423 |
| 15 | Llama 2 13B | 48.81 | #1436 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=toolbench-leaderboard · How It Works · Data refreshed daily, snapshot 2026-10-20.