Pi Agent vs Claude Code After 100 Hours of Real Use

Описание и применение

Pi — минимальный, открытый агент кода, который превосходит Claude Code по количеству пройденных задач и стоимости, но требует ручного управления API и расширений.

Задача

Developers need a lightweight, cost‑effective coding agent that offers full control over tools and pricing, while still achieving high task success rates.

Результат

Pi outperformed Claude Code on a 30‑task benchmark, passing 20/30 tasks (66.7%) versus 16/30 (53.3%) for Claude Code, while costing $0.028 per successful task compared to $0.195 for Claude Code.

Рабочий процесс

• 1. Install Pi from GitHub (MIT licensed).
• 2. Configure desired API keys or run a local model via Ollama/vLLM.
• 3. Launch Pi in the terminal and provide a task prompt.
• 4. Pi calls the model, executes the four built‑in tools (read, write, edit, bash).
• 5. If a missing tool is needed, ask Pi to build a TypeScript extension that runs inside the agent process.
• 6. Use /tree to navigate or branch sessions, and /fork to start new sessions from any point.
• 7. Monitor token usage and cost per task; adjust API provider or model as needed.

Доказательства

Pi passed more tasks than any of the eight harnesses we tested, and did it at the lowest cost of the whole bunch.

Источник

Pi is the cheaper, more flexible, and more transparent harness, especially if you pay per token, switch models, or want full control.

Источник