pi 0.5 (π0.5) API pricing: the waitlist, the one rate card, and the real cost per hour

Pi 0.5 API pricing has no first-party number. The only official API is the Physical Intelligence beta, waitlisted since 13 October 2026. Its post and its form carry no price, checked 26 October 2026. The big marketplaces host none of it. One new third-party platform publishes a rate card row for pi 0.5, read today and dated below. The weights are open, so the other real price is a rented GPU by the hour. This page checks every path, then works the per-call math into a cost per robot hour. The queries behind it, like pi0.5 api pricing and pi 0.5 hosted api, return no priced answer today.

Last verified: 2026-10-26

What you are actually trying to price

Pi 0.5 is a vision-language-action model from Physical Intelligence. It reads camera images, a task sentence, and the joint positions of a robot arm. It returns the joint commands for the next second of motion. The paper came in April 2025. The open weights landed in September 2025 in the openpi repository. The released checkpoint carries 3,616,757,520 parameters. That is about 3.6 billion. Inside it sit a PaliGemma vision-language backbone and a small action expert of about 300 million parameters. The publisher's post says what it is for. It is built for simple tasks done right in rooms the robot has never seen.

One design fact fixes the pricing unit. A call does not return tokens. It returns an action chunk of up to 50 joint commands, built by 10 denoising passes. That is about one second of motion. The publisher calls it a 50-step, one-second chunk. Any API price for this model is therefore a price per call. And a call is worth a different amount of motion on each robot.

The name has four spellings and one collision. People write pi 0.5, pi0.5, or π0.5. In code and config files it appears as pi05. And search results mix in Raspberry Pi boards. Those are small computers with names like Pi Zero and Pi 5. The board cannot run the model. The model needs an NVIDIA GPU with more than 8 GB of memory. The full capability record, benchmarks, and checkpoint guide live on the model's own page: pi05.

The first-party path: the Pi API beta

The only official API from the model's owner opened on 13 October 2026. The announcement describes a full loop, not a simple endpoint. You upload your robot data. You launch fine-tuning on compute that Physical Intelligence manages. You then run the tuned model on your robot through their hosted inference gateway. Their partner examples ran this way. The post is honest about rough edges, including network trouble at some sites.

Here is the pricing content of both pages, read in full on 26 October 2026: there is none. The post carries no rate, no billing unit, and no free-tier terms. The application form carries none either. The form says the team is starting small and onboarding in batches. It asks who would train the policy and what they have done before. Its checklist names ACT, diffusion policy, and pi 0.5. It asks how many hours of teleoperation data you have. Then comes the robot itself: how many you run, what it looks like, and where inference must run. The choices for inference are cloud, local on your own hardware, both, or not sure yet. Nowhere does it ask what you would pay. Nowhere does it state a price.

Two honesty notes on what the beta actually is. First, the post does not name the hosted model. It points to pi 0 and pi 0.5 as the foundation models behind the effort. It then says the API serves their latest and most capable models. The newer releases are those: pi 0.6 shipped in November 2025, pi 0.7 in April 2026, and neither has released weights. A buyer on the waitlist is buying the family line, not a named pi 0.5 endpoint. Second, access is an application, not a checkout. Batch onboarding means no self-serve spend today. The full walkthrough of the form and the access timeline is its own guide: the Physical Intelligence API waitlist.

The third-party state, checked today

On 25 October 2026, a full check of every likely seller found no third-party host for pi 0.5 anywhere. Today that reading no longer holds in full. Here is what each place offers, read on 26 October 2026.

The one priced row, in full, from the abstract.inc pricing page, read 26 October 2026. Their card meters input at 0.15 per million tokens. Cached input is 0.015 per million. Output is 3.70 per million actions. Figures are in US dollars on their page. For pi0.5 they publish an hour estimate of 0.85. Their stated basis is one hour of a 30 Hz control loop, 108,000 actions. Their overview page states the models run on their GPUs and answer in chunks, billed per token and action. That is the only third-party price for this model we found today. It is days old, so treat it as a snapshot with a date, like every rate on this page.

The second new platform is not priced but is real. Robotic Inference covers pi 0.5 in one API alongside GR00T, X-VLA, SmolVLA, and OpenVLA. It adds a leaderboard and machine-readable model contracts. Access runs by invitation while they onboard in batches. Real-robot runs go through a pilot program. No public rate appears on any page we could read today.

The lesson in this table is the dates. The classic marketplaces still carry nothing. Two young platforms cracked that open in October 2026, one with a rate card, one without. Any sentence on the web that says nobody hosts this model is a sentence with a shelf life. Check the date on the claim before you trust it, including this one.

The open-weights path: what self-hosting actually costs

With no priced first-party endpoint, the open weights are the path you can buy today. The repository sets the memory floors. Read them as of 26 October 2026:

The cliff is real. Every 24 GB card serves inference and starts a LoRA fine-tune, which needs more than 22.5 GB. A full fine-tune needs more than 70 GB, so that one rents the 80 GB tier. Hourly rates for the 24 GB tier, named party, read 26 October 2026 on the RunPod pricing page. The RTX 3090 sits at 0.5 per hour. The RTX 4090 sits at 0.89. The RTX A5000 sits at 0.27. Rates move. Treat any rate as a snapshot with a date.

To turn hours into calls you need one more input. You need the time a single call takes on the card. The pi 0 paper publishes a timing table for the same chunk design. The model's own page carries the figures. On an RTX 4090, one chunk takes 73 milliseconds on board. With a network hop, 86. With a per-call time in hand, the raw card cost per call is:

GPU cost per call  =  R x t / 3600

R: hourly rate for the card, from the table above
t: seconds per call, from a benchmark you trust for your setup

One sensitivity line. At the 4090 rate of 0.89 per hour and the 86 millisecond figure, raw card time is about 0.000021 per call. That is a floor at full duty, not a bill. Idle time, restarts, and your own engineering sit on top. The deeper local-versus-hosted treatment, with amortization, is the VLA break-even guide.

The license position, stated in three parts, all read today. The openpi repository code is Apache-2.0. The LeRobot port card carries a Gemma license tag in its own frontmatter. The PaliGemma tokenizer sits behind a license gate on the Hub; you accept its terms and log in before your first run. None of this blocks commercial testing. A buyer should read the repo license and the port card before shipping a product on the weights.

Per call, per action, per hour: the units

Robot API pricing arrives in three units, and most confusion in api cost searches is a unit error. This platform's action lane bills per call. The new third-party card bills per million tokens in and per million actions out. GPU vendors bill per hour. None of these compare directly until you convert.

The conversion starts with the chunk. Pi 0.5 emits up to 50 actions per call, fixed by its training setup, not a request knob. Your control loop decides how often it calls:

calls per robot hour  =  3600 x frame rate / chunk length

The first row is the SO-arm convention used throughout this site: 2,160 calls and 108,000 actions per robot hour. It is the same 108,000-action hour that the abstract.inc estimate is computed on, which is why their 0.85 and any per-call figure can be set side by side at all. Divide any hourly quote by 2,160 and you recover the per-call rate. The reverse works too. That two-way conversion is what makes an hourly figure checkable instead of magical. The full treatment of both units, with worked cases for every meter on the market, is the pricing units guide.

The platform's written record

This platform keeps a written record for the pi 0.5 family on its action route. There is one row each for pi05, pi05-droid, and pi05-libero. Those are the base and the two public checkpoint tunes. The facts of that record, as of 26 October 2026: the rows sit in the platform's model data. Metering is per call. The rate is set. The models are not on the public list, and no date is attached to that. Only listed rows render in the table below, and the models index is the live catalog. This is the same state the model's own page records, in the same words. A written record is not a launch.

Budgeting that record is arithmetic. It takes three named inputs:

monthly spend  =  Rh x C x P
worst case     =  the prepaid wallet balance W

Rh: robot hours per month
C:  calls per robot hour, from the chunk table above
P:  per-call rate, read off the live table for a listed row
W:  wallet balance, the hard stop on spend

The numeric form, at the SO-arm convention. The derivation is published separately and holds the hardware side. At 2,160 calls per robot hour, the written-record rate for the pi05 model id yields 6.48 per robot hour. Work it forward from there. A robot working ten hours a month spends about 65. A 160-hour work month runs about 1,040. Halve the chunk and you double the bill at any fixed per-call rate. Double the chunk and you halve it. The full derivation, with the measured card economics behind it, is the robot inference cost report.

The wallet floor is the point of the shape. An unattended robot loop cannot open a credit line. A refused call bills nothing. The general argument for prepaid floors under robot and agent workloads, with documented runaway cases, is the API bill shock guide.

What to do now

1. Apply to the beta if the hosted path fits. The form facts above are the whole application: team experience, data hours, robot setup, and where inference runs. Expect batches, not instant access, and expect to negotiate price in writing, because nothing is published.
2. Budget the self-host tier now. Serving and a LoRA tune fit the 24 GB rental tier at 0.27 to 0.89 per hour today. A full fine-tune does not; it needs more than 70 GB. Price both before you pick a path.
3. Write your client against a published contract. This platform's action route documents every field: POST /v1/action, multipart form data in, three fields back, with the call count and cost returned on every answer. The run-action reference has the full field list. Your driver can be finished and tested before any spend begins.
4. Budget metered calls from the live table. Listed rates render on the pricing page. You can sign up and create a key without a call from us, and the prepaid balance is the spend stop.

Frequently asked questions

Is there a pi 0.5 API?

Not a priced first-party one. Physical Intelligence runs the Pi API beta, open since 13 October 2026. It is waitlisted and unpriced. Both its post and its form were checked 26 October 2026. The classic marketplaces, Hugging Face, Replicate, fal, and DeepInfra, host none of the checkpoints. One new platform, abstract.inc, publishes a rate card row for pi0.5. It was read today and is quoted above. Another, roboticinference.com, covers the model in an invite-gated beta with no public rate. The open weights in the openpi repository are free to download and serve yourself.

What does pi 0.5 cost per hour?

Three honest answers with three units. A rented 24 GB GPU, enough to serve, cost 0.27 to 0.89 per hour on RunPod on 26 October 2026. Your setup time sits on top. The abstract.inc card estimates 0.85 for a full 30 Hz hour of 108,000 actions, on their token-and-action meter, read the same day. This platform's written record converts to 6.48 per robot hour at 2,160 calls for the pi05 model id, per the published derivation. Any figure you see elsewhere should state its unit, its party, and its date. Otherwise it is not a budget.

How good is pi 0.5?

The paper's headline is generalization. It cleaned kitchens and made beds in homes never seen in training. The publisher also says plainly that it does not always succeed on the first try. On the PhAIL leaderboard, a community-run benchmark read 26 October 2026, pi 0.5 holds the top model score on the pick-and-place suite. That score is 75.1 percent, plus or minus 4.8. GR00T sits at 72.3. Human teleoperation is the reference at 100.0. On the larger 162-task everyday suite it sits near the bottom at 13.8, plus or minus 4.4. One source, one date, both numbers. The capability record with every benchmark is on the pi05 model page, linked at this page's first mention of the model.

How many parameters does pi 0.5 have?

The released base checkpoint carries 3,616,757,520 parameters. That is about 3.6 billion, counted from the model files. The build pairs a PaliGemma vision-language backbone with a small action expert. The expert has about 300 million parameters and generates the motion. For planning, memory matters more than the count. Serving takes more than 8 GB of GPU memory. A LoRA tune takes more than 22.5 GB. A full tune takes more than 70 GB.

Is pi 0.5 the same thing as Raspberry Pi?

No, and the name collision is constant. The model's name starts with the Greek letter pi, written pi 0.5, pi0.5, or π0.5, and pi05 in code. Raspberry Pi is a line of small board computers with product names like Pi Zero and Pi 5, and those boards own big parts of the pi05 search results. The board cannot run the model. Pi 0.5 needs an NVIDIA GPU with more than 8 GB of memory. If you came for the computer, the maker's own pages cover it. If you came to price the robot model, you are in the right place.

The short version: no priced official API, one dated third-party rate card, open weights on a rented 24 GB card, and a per-call budget you can compute now from your chunk length. Listed rates live on the pricing page. Sign up to create a key and budget metered calls.

Last verified: 2026-10-26