Pi Agent vs Claude Code After 100 Hours of Real Use

Description and use

Pi Agent outperforms Claude Code in task success and cost, offering transparency and control, though it requires users to build extensions themselves.

Problem

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

Outcome

Pi outperformed Claude Code on task success and cost per success while offering greater transparency and control. Claude Code remains attractive for users who prioritize built‑in features and predictable subscription pricing.

Workflow

• 1. Set up Pi with desired model API keys or local model. 2. Run Pi in terminal, optionally enabling read‑only tools via --tools read,grep,find,ls. 3. Submit coding tasks to Pi. 4. Pi calls the model, executes tool calls, and returns results. 5. Use /tree or /fork to branch or rewind sessions. 6. If additional tools are needed, ask Pi to generate a TypeScript extension and run it inside the agent process. 7. Monitor token usage and cost per task.
• 8. For comparison, run the same tasks on Claude Code with its default harness and observe higher token consumption and cost per success.

Evidence

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

Source

Pi achieved higher task success (20/30 vs 16/30) at a fraction of the cost ($0.028 per success vs $0.195).

Source