GR00T API pricing: what N1, N1.5 and N1.7 cost to run

No host publishes a per-call price for any GR00T checkpoint, anywhere, as of 26 October 2026. Not NVIDIA, not the robot-only hosts, not the big model hosts. The costs that do exist are hardware costs. One GPU with 16 GB of memory runs inference. A rented H100 works out to about 17.45 per million action steps at the rate Spheron's own guide prints. This page shows the dated checks behind that absence, turns NVIDIA's published speed numbers into a budget you can compute, and explains why the one public robot rate card that does exist does not answer this question.

Last verified: 2026-10-26

What you are actually trying to price

GR00T is NVIDIA's family of vision-language-action models for humanoid robots. A call to one of them does not return text. It returns an action chunk: a short run of motor commands for the next fraction of a second, computed in one pass. That single fact settles the pricing unit before any rate exists. A GR00T price can only be a price per call, a price per action step, or a price per GPU hour. Nobody has published the first two for any GR00T checkpoint.

Four releases carry the name. GR00T N1 arrived on 17 March 2025 with 2.19 billion parameters. GR00T N1.5 followed on 11 June 2025 with 2.72 billion. GR00T N1.6 landed on 15 December 2025 with 3.29 billion. GR00T N1.7 arrived in early access on 17 April 2026 with 3.14 billion. The GR00T version chooser walks every published number version by version. Download counts say which one people pull. On 26 October 2026, N1.7 had 72,458 downloads in the last 30 days. N1 had 426.

The money query behind this page has three search strings: "gr00t api pricing", "nvidia gr00t api cost", and "gr00t n1.5 api". The autocomplete strings under the same intent are about adaptation rather than calls: "fine tuning gr00t" and "gr00t finetuning". That second half matters for the price, because a fine-tune changes the hardware you need and the license you must read. The license guide works through the second one.

The answer, surface by surface

Here is the state on 26 October 2026, checked surface by surface. No host publishes a per-call price for any GR00T checkpoint. No host publishes a per-action-step price for any GR00T checkpoint. The money that exists for GR00T is the cost of the card you run it on.

Two honest notes about that table. First, the absence is dated and every row carries the date. A seller can appear tomorrow, so the table is the part of this page that goes stale first. Second, the empty Hugging Face row is not a guess. It is a machine field. The provider mapping on each NVIDIA GR00T repository is an empty list, and a server-side search across the whole of Hugging Face for a GR00T model with a provider mapping returns an empty result.

The full census of the big hosts, across the whole robot model class rather than GR00T alone, is in the big hosts guide. The absence of a base-model endpoint for this family, with the fine-tune detail, is on the GR00T hosted availability page.

The one public robot rate card, and why it is not GR00T

There are two public robot rate cards in the world. One is Google's, for Gemini Robotics ER 2, and it meters tokens. It is written up in full in the Gemini Robotics pricing guide. The other belongs to a specialist host called abstract.inc, and it meters tokens in and actions out. Its full table lives in that same guide, so this section does not repeat it. What matters here is one row on it.

The nearest name on that card is a trap. The sixth row is called isaac-0.5. It is not NVIDIA's Isaac GR00T. The host's own model documentation names the model id perceptron/isaac-0.5:yam, and describes it as Perceptron's Isaac 0.5 for the bimanual YAM. Every served model on that card drives that same YAM robot. GR00T does not appear on the list at all. So the one public robot rate card that answers this query does not answer it, and the row whose name looks closest belongs to another company.

Two things on that card are worth copying into any rate card for this family. Its output column is metered per million actions, not per token, because a robot action is not a token. And its hour column states its own basis: one hour of a 30 Hz control loop, which is 108,000 actions. A seller of GR00T inference would have to choose from the same small set of units.

What a GR00T call costs when you rent the card

This is the only published GR00T cost math in the world, and part of it has aged. Spheron publishes a GR00T N1 deployment guide dated 3 May 2026. Its cost table is titled "Inference Cost per Million Action Steps". Every money figure in this section is in US dollars.

The guide's own rates at the time of writing: H100 PCIe on demand at 2.01 an hour, H100 SXM5 on spot at 0.80 an hour. Its fine-tuning budget sits next to them: a 20,000-demonstration LoRA run on two H100 SXM5 spot instances, under 20 at those May 2026 rates.

Now the part that has aged. The same page carries a live-rate line. It reads 2.64 an hour on demand, as of 27 October 2026. Multiply the two numbers and the printed figure moves. At 32 Hz an hour carries 115,200 steps, so a million steps take 8.68 hours of the card. At 2.01 an hour that is 17.45, which is the figure in their table. At 2.64 an hour it is about 22.9, roughly a third more than the printed figure. The table is not wrong about the math. It is a May 2026 snapshot of a rate that has since moved.

The guide also carries its own license reminder, in its own words: the GR00T N1 weights remain non-commercial research only. That reminder and the license guide are the reason a GR00T rate card would be a harder sell than it looks. Both numbers in the cost table describe the card. Neither describes the model.

The published speed numbers, and what they cost you

NVIDIA's own N1.7 model card carries an inference timing table for nine accelerators. The table is titled "GR00T N1.7 Inference Timing (4 denoising steps, 1 camera)". Its frequency column is what a budget needs, because a robot control loop consumes action steps at a fixed rate.

Two facts sit under that table. The Isaac GR00T repository states one hardware floor for inference: one GPU with 16 GB of memory or more, with an RTX 4090, L40, H100, Jetson AGX Thor, Orin, or DGX Spark named as examples. Fine-tuning is the expensive phase, and it wants 40 GB of memory or more per GPU, with H100 or L40 nodes recommended. The card is the same price. The motion is not: at one hourly rate, an H100 running at 35.9 steps a second delivers 129,240 steps in an hour, and an AGX Thor running at 10.7 delivers 38,520.

The cost arithmetic from those two tables is one line each:

cost per action step  =  R x t / 3600
cost per hour of motion  =  R x 3600 / f

R: hourly rate for the card
t: seconds per action step
f: action steps per second, from the published table

Hold the hourly rate fixed and read the two rows against each other. The same money buys three times the motion on the data centre card that it buys on the edge module. That is the whole economic case for putting the model near the robot, and the Jetson versus cloud guide works the trade.

What a metered per-call price looks like

No host prices GR00T today, so here is the shape you would buy into when one does. This platform's action route is POST /v1/action. It takes multipart form data and returns JSON. The request fields are model, instruction, state, and one or more image parts. The response carries three fields: actions, normalization, and usage. One served request bills one call, and a refused request bills nothing. The refusals have stable codes. A field the route does not take comes back as 400 unknown_param, naming the field. An empty prepaid wallet comes back as 402 credit_balance_exhausted. A model id that is not on the public list comes back as 404 model_not_found. Every response carries an x-request-id header. The field-by-field contract is on the action route page, and the generated listing is in the run-action reference.

The platform keeps a written record for the pi05 model id and for molmoact2-so100-101. As of 26 October 2026 both records carry a rate, both are metered per call, and neither model id is on the public list. The pi05 model page and the molmoact2 page say the same in their own words. That record carries no promise. A written record is not a launch.

Below is the platform's live rate card. It is generated from the platform's model data at render time, so the numbers in it are current when you read this page and not when it was written. The pricing page is the standing source for the same numbers.

Budgeting a metered robot lane is arithmetic on 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
P:  per-call rate, read off the live table for a listed row
W:  wallet balance, the hard stop on spend

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 agent and robot workloads, with documented runaway cases, is in the API bill shock guide. The deeper treatment of per-call versus per-robot-hour units, with worked cases, is in the pricing units guide.

What to do now

1. If you need GR00T, self-host it. Rent the card, pull the checkpoint from nvidia/GR00T-N1.7-3B, and follow the authors' own inference script. Your cost is the hourly rate plus your setup time. The NVIDIA repository and the hardware floor above tell you which card to pick.
2. If you need a metered action API today, pick a model that has one. The action route, the field names, the refusal codes and the wallet are all published now, so a driver can be written and tested before any spend begins. The models index is the live catalog. You can sign up and create a key without talking to anyone.
3. If you are fine-tuning, budget the card and the licence, not the API. A fine-tune wants 40 GB of memory or more per GPU, and the weights carry a research and evaluation limit. Both are read in the version chooser and the license guide.
4. If you are comparing the class, start from the one priced card. The Gemini Robotics pricing guide has the only two public robot rate cards side by side, with the unit conversions worked out.

Frequently asked questions

Is there a GR00T API?

Not one you can buy. On 26 October 2026 no marketplace sells hosted access to any GR00T checkpoint. NVIDIA's build catalogue carries no GR00T model. The four NVIDIA repositories on Hugging Face carry an empty hosted provider mapping each. Together AI and Fireworks carry no GR00T entry. The specialist robot host lists four GR00T N1.7 fine-tunes with a sleeping status and publishes no rate card. If a seller appears, the absence table on this page is the first part to go stale.

What does GR00T cost per hour?

Compute, not inference. NVIDIA's own floor is one GPU with 16 GB of memory or more. The only published GR00T cost table in the world is Spheron's, and it prices card hours: 17.45 per million action steps on an H100 PCIe at its May 2026 rate of 2.01 an hour. The same page's live-rate line reads 2.64 an hour as of 27 October 2026, which puts the same line closer to 22.9. Both numbers buy the card. Neither buys the model.

Which GR00T version should I price?

N1.7 for anything new. It is the current release, it carries the published speed table above, and it pulls about 46 times the monthly downloads of N1.5. N1.5 only makes sense for an experiment that already started on it. N1 only makes sense for history. The version chooser has the full table.

Can GR00T be used in a commercial product?

Read the licence file, because the claim layer and the file disagree. The licence file inside every GR00T release, N1.7 included, permits research and evaluation only. The N1.7 model card says the model is ready for commercial and non-commercial use, and NVIDIA's April 2026 launch post says commercial licensing is now included. The file has not changed. This page is engineering guidance and not legal advice. The licence guide puts all four claims side by side.

Why does no host sell GR00T?

Three reasons stack. The licence makes a resale story hard to write. The model needs a GPU with 16 GB of memory or more, and a per-call price has to cover an idle card as well as a busy one. And the demand that would fill a card is still research demand: the search strings behind this query are about fine-tuning, and a fine-tune needs many hours of a good card, not a cheap call. That is also the shape of the market, with the per-robot-hour glossary on the unit and the measured robot-hour report on the numbers.