Physical Intelligence2026-10-14 · Joe Zhou
The Pi API Beta Is a Four-Day Loop. The Demo Is Not the Rate.
Physical Intelligence opened the Pi API beta on 13 October 2026. Upload data, fine-tune on their managed infrastructure, run inference through a hosted gateway. The named early result is Almond’s Axol deburring spacers after four days of iteration. Most shown applications used cloud inference, and the page flags network connectivity as a challenge for some.
Physical Intelligence published the Pi API beta on 13 October 2026. The workflow is upload data, launch fine-tuning on compute and infrastructure they manage, and access the fine-tuned model via a hosted gateway for inference. The page says the latest model available through the API already provides more data-efficient fine-tuning and improved capability for their partners. They plan to keep updating the models that sit under the API.
The open weights remain on the openpi repository. The API is the managed layer on top.
The early access table
The page lists partners and early-access users. These are the names they chose to show.
| Company | What the page shows | Named detail | Source |
|---|---|---|---|
| Actor | Autonomous wheel loader scooping material into a dump truck | Real commercial deployment, no operator inside the machine | Physical Intelligence blog, 13 Oct 2026 |
| Almond | Axol dual-arm robot deburring spacers used in its own assembly | Model obtained after iterating with the Pi API for four days | Physical Intelligence blog, 13 Oct 2026 |
| AgRobotics | Listed among broad-application partners | Harvesting context in secondary reports | Physical Intelligence blog, 13 Oct 2026 |
| Chef Robotics / CloudChef / Omakase | Food preparation and plating | Listed in early-access set | Physical Intelligence blog, 13 Oct 2026 |
| Formic Robotics | Listed in early-access set | Formic has previously described robots in factories | Physical Intelligence blog, 13 Oct 2026 |
| General Instinct | Listed in early-access set | Secondary reports say three days of access and matching an adapted open model with 5 percent of that model's data | Physical Intelligence blog, 13 Oct 2026; secondary coverage |
| Ultra, Weave, others | Listed | Deployments from Japan to Switzerland to the United States | Physical Intelligence blog, 13 Oct 2026 |
Most applications shown ran with cloud-hosted inference through the API. The page says these early deployments highlighted areas for improvement, including network connectivity challenges for some.
What the selected demos hide
Four days is a calendar claim for one dual-arm deburring task on one robot. It is not a success rate, a data volume, a cycle time, or a failure count. The page does not say how many demonstrations were uploaded, how many fine-tunes were launched, or what fraction of attempts succeeded on the hardware.
The commercial wheel-loader clip is presented as a real deployment. The page does not give hours, tons, or intervention rate for that machine.
Secondary coverage of General Instinct reports training within three days and matching an existing adapted open model with 5 percent of that model's adaptation data. That number sits in the secondary write-up, not on the Physical Intelligence page.
What is not in the record
We do not have a success rate, an intervention count, or a data volume for any of the shown tasks.
We do not have a latency number for the cloud inference path, or a count of how many of the shown applications hit connectivity problems.
We do not have a price, a compute rate, or a data-retention policy for the beta.
We do not have the waitlist acceptance criteria or a timeline for general access.
We do not have an independent measurement of the "more data-efficient fine-tuning" claim.
The sentence
The Pi API beta is a managed iteration loop on a waitlist. The four-day Almond result is a calendar claim on one task, not a rate you can take to a second robot.
If you use this note
Quote the sentence you actually used. Link the URL. If you want the referral on the citation, append `?via=yourcode` — mint the code at Take.
"The Pi API beta is a managed iteration loop on a waitlist. The four-day Almond result is a calendar claim on one task, not a rate you can take to a second robot." Source: Roboterminal · https://roboterminal.app/posts/pi-api-beta-four-day-loop