OccuBench (Mixed Faults): leaderboard

OccuBench (2026) evaluates agents on 382 professional tool-use tasks from 100 occupational scenarios in 10 industries, against environments simulated by Gemini-3-Flash-Preview, each trajectory passed or failed by a rubric verifier; thinking mode on, high effort where configurable. This board injects a mix of explicit and implicit faults; completion rate over all tasks in percent; higher is better.

Metric: Completion rate (%) over all 382 tasks with about half explicit and half implicit injected faults (E3, two transient fault events of two calls each); 382 tool-use task instances from 100 professional scenarios, the environment simulated by Gemini-3-Flash-Preview and each trajectory passed or failed by a rubric verifier; thinking mode on, high reasoning effort where configurable; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.2High67#185
2Gemini 3.1 Pro (Preview)65.2#99
3Claude Opus 4.6High63.9#113
4Qwen 3.5 PlusThinking54.2#242
5Claude Sonnet 4.6Thinking52.9#173
6DeepSeek V3.2Thinking51.6#419
7GLM-5Thinking47.4#269
8MiniMax-M2.746.9#539
9Kimi K2.5Thinking40.1#266

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

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