PolyWorkBench (Codex): leaderboard
PolyWorkBench (Beijing Jiaotong University and Tencent, 2026) runs LLM agents through 67 multilingual long-horizon workplace workflows in commerce, knowledge work, legal analysis, localization and manufacturing, scoring the produced artefacts with weighted deterministic rubric checks. Every model runs inside the OpenAI Codex CLI harness. Mean Grade in percent from the best of up to three runs, higher is better.
Metric: Mean Grade (%; weighted deterministic rubric checks of the output artefacts, mean over 67 multilingual long-horizon workplace tasks in five domains; best of up to three runs, Pass@1; the OpenAI Codex CLI in full-access mode, 1,800 s per task). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 88.7 | #138 |
| 2 | GPT-5.5 | 80.8 | #49 |
| 3 | DeepSeek V4 Flash | 79.7 | #258 |
| 4 | GLM-5.1 | 78.1 | #179 |
| 5 | Qwen 3.6 27B | 76.6 | #341 |
| 6 | Claude Opus 4.8 | 69.8 | #56 |
| 7 | Qwen 3.6 35B A3B | 68.2 | #462 |
| 8 | Claude Opus 4.7 | 61.2 | #79 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=polyworkbench-codex · How It Works · Data refreshed daily, snapshot 2026-11-13.