Physical Intelligence API Beta: Frontier Models for Every Robot

Summary

Physical Intelligence (π) opened a waitlisted beta of the Pi API, which lets companies fine-tune its robot foundation models on their own data and run them through hosted inference. No pricing is given, and the newest model behind the API is not named.

Key points

• Workflow: upload robot demonstration data, launch managed fine-tuning, serve the result through a hosted gateway.
• Named models: π0 and π0.5, whose code is public in the openpi repository; the API's latest model is described only as more data-efficient to fine-tune.
• Early partners in heavy machinery, agriculture, food preparation, and manufacturing; one (Almond) reached a working model in four days.
• Most deployments used cloud inference, and some had network connectivity problems.

Why it matters

Robot foundation models are moving to the same fine-tune-and-serve API model as LLMs, on the same day Google pitched Gemini 4 Argon for robotics.