Composio 于 8 月 6 日发布了一项比较研究结果，针对四个 AI 编程智能体框架在 DeepSeek V4 Flash 模型上的表现进行了系统测试[1]。这家总部位于旧金山、成立于 2023 年的公司采用 30 项任务作为评估基准，使用 Gmail、GitHub、Slack 和 Notion 等真实工具进行测试[1]。
测试结果显示，各框架在不同维度的表现差异明显[1]。Oh My Pi 以 17/30 的成功率领先，虽然耗时较长（每任务 272 秒），但完成任务数最多[1]；Claude Code 在速度上最具竞争力，平均每任务耗时 122 秒，但成本最高（每成功任务 0.195 美元）[1]；Codex 完成了 16/30 的任务[1]；OpenCode 虽然成功率最低（14/30），但成本最优，每成功任务仅需 0.073 美元[1]。
Composio released comparative testing results on August 6 showing how four AI programming agent frameworks perform when paired with the DeepSeek V4 Flash model [1]. The San Francisco-based company, founded in 2023, evaluated Oh My Pi, Claude Code, Codex, and OpenCode across 30 real-world tasks involving Gmail, GitHub, Slack, and Notion integrations [1].
Oh My Pi emerged as the most reliable option with a 17 out of 30 success rate, though it required the longest execution time at 272 seconds per task [1]. Claude Code ranked second in success rate with 16 completed tasks but achieved the fastest average speed of 122 seconds per task, though at the highest cost of $0.195 per successful task [1]. Codex also completed 16 tasks, while OpenCode successfully finished 14 tasks but maintained the lowest operational cost at $0.073 per successful task [1].
