ToolBench - WebShop Long: leaderboard
ToolBench WebShop Long evaluates text-only shopping agents that choose search[text] and click[button] actions to satisfy product constraints. The Long split uses full human shopping trajectories as in-context demonstrations; its 1,533 demonstrations often exceed model input limits. Generated actions are executed in WebShop and scored by task reward.
Metric: Task Score. Source: huggingface.co. Saturation forecast: Not forecast. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | starcoder | 2.6 | #1577 |
| 2 | Llama 2 70B | 1.53 | #1201 |
| 3 | Llama 2 13B | 0.6 | #1436 |
| 4 | GPT-3.5 Turbo | 0 | #865 |
| 5 | GPT-4 | 0 | #438 |
| 6 | Llama 2 7B | 0 | #1520 |
| 7 | Dolly V2 12B | 0 | #1620 |
| 8 | pythia-12B | 0 | #1621 |
| 9 | LLaMA-65B | 0 | #1303 |
| 10 | StableLM Tuned Alpha 7B | 0 | #1640 |
| 11 | CodeLlama-7B-Instruct-hf | 0 | #1545 |
| 12 | CodeLlama-34B-hf | 0 | #1515 |
| 13 | opt-1.3B | 0 | #1636 |
| 14 | CodeLlama-13B-Instruct-hf | 0 | #1532 |
| 15 | opt-30B | 0 | #1610 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=toolbench-webshop-long · How It Works · Data refreshed daily, snapshot 2026-10-20.