Pre-launchS1 Nano is in training. Runtime is in private beta.What is not done

Pricing

Pay for the rack, not the model.

The weights are free under Apache 2.0. What you pay for is inference on machines that sit beside a validator, which is the part you cannot easily reproduce. Drag the marker to put your own volume through it.

S1 Nano at 50,000 decisions a day

$140.40/mo

Per decision$0.0001Tokens in and out900 / 90A year$1,685
Drag the marker, or use the arrow keys, to set your volume. Curves come from the published prices and a token budget per decision: 900 in and 90 out for Nano, rising with context. Verify is flat because it is billed per decision. Nothing here is billable yet, so treat it as a planning number rather than a quote.

Per million tokens.

One rate card. No seat licences, no minimum commitment, no enterprise tier that exists mainly to start a conversation.

S1 Nano

Minimum billing unit is 1K tokens.

In
$0.08
Out
$0.24

S1

Batch window at 50 percent.

In
$0.45
Out
$1.40

S1 Pro

Active-parameter pricing. 9B active of 140B.

In
$1.80
Out
$6.00

S1 Verify

Flat. One transaction, one decision.

In
$0.05
Out
per verification

United States dollars.1

What the price includes.

Everything that makes a decision safe to sign. None of it is a paid add-on, because an agent that skips any of it is not a product we want to sell.

Colocated inference
Your calls land on hardware in the same facilities as the validators, not in a general cloud region several hops away.
Simulation on every call
Not billed separately, and not possible to turn off. A plan that has not been replayed against current state never reaches a key.
The policy engine
Value ceilings, slippage limits, allowed programs and counterparties, and enclave signing. Checked on every call at no extra charge.
S1 Verify inline
The refusal step runs before the signer whether or not you asked for it. You are billed for it per decision, at the flat rate above.
Running it yourself
Free. Apache 2.0 weights, the training recipe, the data card and the evaluation harness. You lose the colocation, which is most of the latency argument.

When self-hosting is the right answer.

Often. We would rather say so here than have you work it out after signing something.

You already run a validator
Then you own the one thing we charge for. Runtime installs beside it and the weights are free. Talk to us about the policy engine on its own.
Your volume is steady and large
Past a few million decisions a day, a reserved box beats per-token pricing. The cost model above crosses over visibly.
You cannot send intent off your own network
A plain-language intent string is a sensitive thing to hand to a vendor. If that is a problem, self-host; it is the honest recommendation.
You need S1 Pro occasionally
Pro is priced on active parameters, 9B of 140B, because that is what it costs us to serve. For bursty research use, hosted is almost always cheaper than idle hardware.

Nothing is billable yet.

No model has shipped, so no account can be charged and no card can be taken. These prices are published early so that anyone costing out a build has a real number to work with rather than a sales call.

Talk to us
  1. Published prices are an intention, not an offer. They may move before the first model ships, and nothing on this page can be purchased today. The cost model is built from these rates and a token budget per decision; your own prompt sizes will differ. What is not done.