北京大学与元空AI联合实验室正式开源了科研智能体项目OpenAI4S，这是对Anthropic闭源产品Claude Science的开源复现 [1]。该项目于2026年7月6日在GitHub发布，采用MIT许可证 [1]，核心零依赖，内置30多项科研技能 [1]。
OpenAI4S的核心特性是Code-as-Action（代码即行动），支持代码执行、数据分析、GPU算力接入和版本化科研产物 [1]。项目内置的科研能力包括14个需要GPU的科学计算技能 [1]，并采用不伪造策略，拒绝模拟数据和虚假结果 [1]。项目支持用户接入自有计算环境，可将GPU任务派发到用户服务器 [1]。该举措旨在让AI真正进入科研工作流，而非仅停留在演示阶段 [1]。项目的主要贡献者包括张恭博、李昊、王豫、刘宇阳等 [1]。
Beijing University and Yuan Kong AI Joint Laboratory have officially open-sourced OpenAI4S, a scientific research agent project that reimplements Anthropic's proprietary Claude Science offering [1]. Released on GitHub on July 6, 2026, the project operates under the MIT license and features a core architecture with zero dependencies [1].
The initiative incorporates over 30 built-in research skills designed to integrate artificial intelligence directly into scientific workflows rather than limiting it to demonstration purposes [1]. OpenAI4S implements a Code-as-Action framework and includes 14 capabilities requiring GPU-based scientific computation [1]. The project adheres to a no-fabrication policy, explicitly rejecting simulated data and false results to ensure research integrity [1]. Additionally, the platform supports integration with users' own computing environments and enables dispatch of GPU tasks to user servers [1].
