WildToolBench: leaderboard

WildToolBench (2026) evaluates LLM tool use under realistic user behavior: 256 multi-turn scenarios over about 1,600 public APIs mix compositional tool orchestration, intentions hidden in earlier turns and switches between tool, chat and clarification requests. This board is session accuracy in percent, the share of dialogues in which all four user tasks are completed correctly, each model calling tools through its native function-call format; higher is better.

Metric: Session accuracy (%): share of the 256 dialogues in which all four tasks are completed correctly, on WildToolBench (256 multi-turn scenarios, four user tasks each, 1,024 tasks, built from real user-log patterns with expert-annotated tool calls), each model through its native function-call format with default decoding; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 50 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.0 Flash (Thinking)14.45
2Gemini 2.5 Pro14.06#276
3Claude Sonnet 412.5#410
4O112.11#403
5GLM-4.512.11#542
6GPT-4o11.72#751
7Claude 3.7 Sonnet11.33#470
8Kimi K210.55#443
9Grok 410.16#319
10O310.16#237
11Qwen 3 30B A3BNon-reasoning9.77#1156
12DeepSeek R19.38#482
13DeepSeek V39.38#729
14Qwen 3 14BReasoning9.38#1269
15Claude Opus 4.19.38#278

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Example tasks

这个周末我想去外地游玩,请帮我查下周末两天Chicago的天气。

Source

Interactive version: theaggregate.ai/benchmark?slug=wildtoolbench · How It Works · Data refreshed daily, snapshot 2026-11-21.