Yuetai Li et al.

JobBench

An occupational agent benchmark for professional workflows that workers say they most want delegated to AI.

In index
Agentic
Best in agentic
Source ↗

Models36

Top result75.3%Muse Wonder Full

Top-3 spread4.8 pts

Year2026

Result and price

32 of 36 results

Results

36 results

#ModelResult

01Muse Wonder Full75.3%

02Muse Wonder70.5%

02Inkling 2 Small70.5%

04LongCat-2.565.4%

05Muse Spark 1.364.9%

06Ultramarine 1 Flash62.9%

07MiMo-V2.6-Pro62.0%

08Muse Spark Open61.8%

09Hy4 preview61.7%

10MiMo-V2.6-Flash61.2%

11Step 5 Preview59.0%

12Qwen3.8-Flash-Next55.7%

13Muse Spark 1.154.7%

14Qwen3.8 Max53.4%

15Kimi K352.9%

16Atria Dawn Preview50.3%

17Claude Opus 4.7 (Adaptive)45.9%

18GPT-5.542.7%

19GPT-5.438.9%

20Claude Sonnet 4.636.9%

21Claude Opus 4.636.7%

22GPT-5.234.3%

23GPT-5.3 Codex33.7%

24Qwen3.8-27B33.4%

25Claude Opus 4.532.3%

26Claude Sonnet 4.527.7%

27GPT-5.1-Codex26.2%

28GPT-5.2-Codex26.0%

29Claude 4.1 Opus21.9%

30Qwen3.5 Plus18.5%

31Claude 4 Sonnet18.4%

32Claude Haiku 4.516.0%

33Gemini 3 Pro11.4%

33Gemini 3 Flash11.4%

35Kimi K2.58.7%

36GPT-5 (high)8.5%