Skip to main content
An agent that needs training data can buy it outright: no signup, no dashboard, no human. You describe the behaviour you want, pay a stated price in USDC, and collect a validated JSONL file a few minutes later. Everything here is one endpoint and one status URL.

​
Why it is not a single call

Generating a dataset takes minutes, and a large one can take considerably longer. Nothing holds an HTTP connection for that: Cloudflare and most agent HTTP clients give up around 100 seconds. So payment and manufacturing are separate. Paying returns in about a second with a dataset_id and a claim_token. Generation runs in the background and you either poll for it or receive a signed webhook.

​
1. Check the price

Price is (generation + verification) x rows / 100. Every dataset is checked by default, and that check is its own line, so 200 examples on fast costs 0.60,200onthegridcosts0.60, 200 on the grid costs 0.36, and 2,000 on best costs $22. Send "verify": false to skip checking and pay for generation only. The per-million-token figures are the provider’s published rates, shown so you can compare models. You pay the flat per-example price, not per token.

​
Every row is checked

A second model reads each generated row back and scores it against your rules before it is allowed into the dataset. Rows that break a rule are thrown away and regenerated, so you receive a full-size dataset of rows that passed rather than a full-size dataset with failures in it. Who does the checking depends on where you generate, and the difference matters: Managed jobs are checked by a separate frontier model, never the one that wrote the rows, because a model marks its own homework badly. That costs $0.10 per 100 examples. Grid jobs are checked by the grid itself, using the network’s own models, so your data never leaves the confidential network even to be verified. That costs $0.03 per 100. The honest trade is that on-network models are smaller, so it is a lighter check than the frontier judge, which is why it costs a third as much. The status response carries the receipt, so you can see what happened: checked, rejected, and unchecked if the budget ran out before a batch could be judged. A job never claims a check it did not perform.
This is why generated data beats prompting a model yourself in a loop. A real 200-row support dataset built without rules or checking invented twelve different compensation policies, including thirteen full refunds, because nothing ever verified the output against a policy.
Two ways to generate. Managed models are the fastest with the strongest data quality. The decentralized grid runs on our own confidential network, end to end encrypted, where nothing is sent to an outside provider: it costs less and takes longer, and if a node drops mid-job the next most appropriate model on the network takes over automatically.

​
2. Ask, and get a price

POST /datasets/x402/synth with no payment header returns 402 with a standard x402 challenge plus a singularity block describing what you are buying.
Between five and twenty seed examples are required: they are what the generator learns your format and tone from. Sizes run from 50 to 2,000 rows. Set house rules. Pass "rules": ["Never offer more than a 15% discount", "Never promise a refund"] (up to 12, 200 characters each) and every generated reply must obey them. A standing instruction says what the assistant does; rules say what it may never do. Without them the model invents its own policy and applies it differently in every row. In one real 200-row support set it handed out twelve different compensation offers, including full refunds. Consistency of policies, figures and promises is enforced by default even when you set no rules.
Quotes last ten minutes. The price is snapshotted when it is issued, so a rate change can never alter what you are charged.

​
3. Pay

Sign the challenge and send it back with the same body plus the quote_id.
Save the claim_token. It is shown once and it is the only way to read the job or download the result. It is stored only as a hash, so it cannot be recovered.
The payment is bound to the exact job you were quoted. Sending a payment with a different body is refused, so a captured payment cannot be redirected at a different (or larger) dataset.

​
4. Collect it

While it runs:
When it is done:
The download link is presigned and lasts five minutes; poll again for a fresh one whenever you need it. The file is JSONL, one conversation per line, ready for fine-tuning as-is. If generation cannot produce enough usable examples the job fails and the fee comes back to your wallet as credits:
refund_status is read from the ledger rather than asserted, so pending genuinely means the refund is still being applied.

​
Webhooks

Pass a webhook_url when you ask for the quote and we will POST to it when the job reaches a terminal state, so you do not have to poll at all.
Every delivery is signed with the webhook_secret returned in the 202:
Verify it by computing HMAC-SHA256(secret, "<t>.<raw body>") and comparing to v1, then rejecting anything where t is more than a few minutes old. That timestamp is what makes a captured delivery non-replayable. Redirects are not followed, and only a 2xx counts as delivered; we retry three times (one minute, five minutes, twenty-five minutes) and then stop. Delivery never affects the job, so polling always works as a fallback.

​
Chains

USDC on Base and Solana, USDG on Robinhood Chain, USDM on MegaETH. Pick with "network" in the request body; the challenge comes back with the right asset and amount for that chain.

​
Through MCP

If your agent speaks MCP, the same flow is four tools on https://mcp.x402layer.cc/mcp:

​
What you actually get

Not raw model output. Every line is schema-checked and dropped if malformed; near-duplicates and reworded repeats of an earlier scenario are rejected rather than padding the count; the generator is steered across a rotating set of coverage axes so you get edge cases instead of fifty variations of one; and every row carries the same standing instruction, so the file is uniform. If we cannot reach a usable floor, you get your money back.