Alibaba's Qianwen Office Tops Wall Street Test of 8 AI Agents
Jefferies analysts tested eight AI agents on real office tasks, and Alibaba's Qianwen Office ranked first overall, beating Claude Cowork and Codex.
The test covered five real office tasks: summarizing a company's annual report from multiple documents, searching and comparing company operating data online, operating a real desktop browser for information retrieval and document generation, creating an English PPT based on data, and generating a marketing poster from a reference image. Qianwen Office performed consistently well, standing out in complex office tasks, browser control, and multimodal content generation. It was the only agent product to score above 90 in all evaluation dimensions.
The report further broke down agent capabilities into model and harness. The harness refers to the engineering mechanisms around the model, including instructions, context, tool invocation, boundary control, feedback correction, and governance. According to Jefferies' calculations, Qianwen Office's implied harness score ranked first among the tested products, higher than Claude Cowork, Codex, and five other mainstream domestic and international agents. This indicates that beyond the underlying model, how to convert model intelligence into stable task results through engineering and product capabilities is becoming an important dimension of agent competition.
Cost is also a key factor in agent commercialization. The report showed that the API price of Qwen 3.8 Max, the underlying model of Qianwen Office, is significantly lower than some leading overseas models. Since agents typically require multi-round reasoning, continuous tool calls, and long-chain task execution, enterprises may increasingly pay attention to the cost per task when evaluating agent value. The combination of Alibaba's Qwen model and Qianwen Office agent demonstrates a solid performance-to-cost advantage.
In recent months, agent competition has been shifting from personal office use to enterprise-level scenarios. The report believes that for enterprise-grade agents, workflow and ecosystem collaboration are core competitive barriers. As agents continuously connect to corporate data, business systems, collaboration tools, and permission systems, users' historical tasks, work habits, connectors, skills, and automated processes will accumulate in the product, creating higher stickiness and switching costs.
Qianwen Office is also designed for enterprise AI needs. It has been initially integrated with DingTalk IM, allowing employees to complete group chat summaries, document and spreadsheet creation, and message and email sending directly through Qianwen Office. In the future, enterprise customers will be able to connect Qianwen Office to real workflow systems, linking corporate databases and workflows to improve office and organizational efficiency.