Physical Intelligence Pi API beta: a week in, still waitlist-only with no public price

As of 8:55 a.m. ET, Tuesday, Oct. 20, 2026

A week after Physical Intelligence opened a beta of its Pi API on Oct. 13, access is still by application, and the company has posted no price on its announcement or waitlist pages. The company says the service lets robot developers fine-tune its latest models on their own data and run them through hosted inference.

The company's blog and its X account show no new posts since the Oct. 13 announcement, as of Tuesday morning.

What developers can do

"You upload data, launch fine-tuning on compute and infrastructure we manage, and access fine-tuned models via a hosted gateway to run inference," Physical Intelligence wrote in its Oct. 13 post.

The company says the latest model available through the API "already provides more data-efficient fine-tuning and improved capability for our partners." The post does not name the model version. It says the company plans to keep giving API users access to its "latest and most capable models."

The company describes the workflow as a loop: collect demonstrations, train a model, test it on the robot and use the failures to decide what data to collect next.

Who is in the beta

Physical Intelligence says it "gave a small number of companies early access to the API." The post shows work from Almond, Argon Robotics, Chef Robotics, Deft, Dream Machines, Formic Robotics, Foundry, Frontier Machines, General Instinct, Intuition.dev, Omakase Robotics, robot.com, Ultra and Yondu, plus a hackathon hosted by Actor.

A separate group is described as companies Physical Intelligence works with closely to build autonomy systems: Actor, AgRobotics, CloudChef, an unnamed dairy partner, Ultra and Weave.

Factories, kitchens and farms, as the company and partners describe them

All of the results below are claims by Physical Intelligence or its partners, shown in the company's post with short videos. GROKsiNEWS.com has not independently tested them.

Factories. Deft, which is building wheeled humanoids for automotive manufacturing floors, "more than doubled the success rate in six days" on a part-loading task, the company says. Foundry uploaded about 500 demonstrations for an engine assembly task and was testing fine-tuned models on its factory floor by the next day, according to the post.

General Instinct trained a server-assembly policy within three days that the company says matched an existing fine-tuned open-source model "while using only 5% of the data" that model was fine-tuned with. Frontier Machines trained its own packaging model within a week that matched one built by Physical Intelligence researchers, then deployed it at a customer site, according to the post.

Kitchens. Chef Robotics trained a two-armed burger-assembly model from 33 demonstrations collected the day before, the company says. Argon Robotics, which builds robots for backend kitchen tasks, had a dish-racking model within 35 hours of getting access and switched its deployment to it the next day, according to the post. CloudChef's robots portion ingredients into containers in commercial kitchens, the post says.

Farms. AgRobotics, based in Salinas, California, builds a strawberry-harvesting robot that works in the furrow, the post says. At a dairy in Stanislaus County, California, an autonomous tractor pushes feed toward cows "without human intervention for hours at a time," according to the company.

Physical Intelligence also shows Actor's wheel loader filling a dump truck in Utah with no operator inside.

The waitlist

"We're starting small and onboarding in batches," the waitlist page says. The form asks who on the team would train the policy, how many hours of teleoperation data they have, how many robots are available and what the robots look like.

It also asks where inference needs to run, offering cloud, local hardware, both or "not sure yet." The form does not say whether local deployment is offered.

What the company has not said

Neither the announcement nor the waitlist page lists prices, usage limits, a service-level agreement or a general-availability date. GROKsiNEWS.com could not find public developer documentation, though the post says one partner worked "from the API docs."

The company flags one limitation itself. Most applications shown ran on cloud-hosted inference, and those early deployments "highlighted areas for improvement, including network connectivity challenges for some," it wrote.

Separately, the company's openpi repository on GitHub, which lists an Apache 2.0 license, offers base model checkpoints for its earlier pi0, pi0-FAST and pi0.5 models that developers can fine-tune on their own data. Neither the repository nor the API announcement says those are the models behind the API.