Hybrid computer-use agents

Explore.
Build. Verify.

Can an agent recreate an application
just by using it?

Real software. Recreated by an agent.macOS
GPT-6 Astra recreation of Food Truck
Original Food Truck on macOS
ReferenceAgent recreation

Drag anywhere on the image to compare. Use arrow keys for small steps, Page Up or Page Down for larger steps, or Home and End to reveal one full image.

Food Truck · GPT-6 AstraFull images: Reference ↗Recreation ↗
250held-out applications
5native platforms
2scoring channels

Ubuntu · macOS · Windows
Android · Web

RecreationBench

How well do agents recreate?

Interested in evaluating self-hosted models or curious about how other models perform? Please contact xiezhihui.xzh@alibaba-inc.com or gaochang.gao@alibaba-inc.com.

RecreationBench model scores, programmatic threshold coverage, and estimated API cost per task. Threshold coverage is the percentage of applications reaching the stated Prog score; Overall averages platform rates equally. Higher scores and coverage are better; lower costs are cheaper. Cost bars share a linear scale across all models. Sorting by cost retains Average score ranks.
Rank
Loading scores…

Paper analysis

What do the results reveal?

Cost and benchmark performance

Loading paper analysis data…

Citation

BibTeX
Download .bib
@misc{qwen2026recreationworld,
      title={RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents},
      author={Shuai Bai and Jiayong Deng and Sicheng Fan and Yikun Fu and Chang Gao and Xuhao Hu and Mianqiu Huang and Yizhen Jiang and Yuheng Jing and Dehui Kong and Keliang Li and Ning Li and Wanli Li and Dayiheng Liu and Dunjie Lu and Changwei Luo and Que Shen and Zheyuan Wang and Zijian Wang and Jie Wu and Gao Wu and Zhihui Xie and Rui Xie and Haiyang Xu and An Yang and Jiakang Yuan and Yanming Zhang and Jiajun Zhang and Xi Zhang and Zhenru Zhang and Zhuo Zhen and Mingkang Zhu and Bowen Zhou},
      year={2026},
      eprint={2609.22000},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.22000},
}

Reference application screenshot · Try the live comparison ↗