x402 Explained: How AI Agents Pay for Things Online

A status code nobody used

The HTTP specification reserved status code 402 for "Payment Required" and then left it empty for about thirty years. Every other code found a job. This one waited for a use case that did not exist yet: a machine that needs to buy something, immediately, for a fraction of a cent, without a person in the loop.

x402 is the standard that finally uses it. Whether the machines doing the buying are actually AI agents is a separate question, and the honest answer is mostly not yet.

How the flow works

The mechanism is deliberately small. There is no new transport, no session and no account.

1. A client requests a resource: an API endpoint, a dataset, an article, a model inference.
2. The server responds 402 Payment Required, with instructions attached: the amount, the asset, the chain and where to send it.
3. The client signs a stablecoin payment for that amount and repeats the request with the proof attached.
4. The server verifies the payment and returns the content.

The entire cycle completes in seconds, and nothing about it requires the client to have registered, stored a card, or been a human being. For a server, it converts an audience of subscribers into an audience of anyone who can pay 30 cents, including software.

Most implementations put a facilitator between the server and the chain to verify and settle, so the origin server does not have to run blockchain infrastructure to accept a payment.

Who is behind it

x402 originated at Coinbase, and a foundation formed with Cloudflare in 2025 to push it as a shared standard rather than a single company's protocol. Stripe, Google and Cloudflare have all built support into their own services, which is the part that matters: a payment standard with one implementer is a product, and a payment standard with several is infrastructure.

The specification itself is public, and Coinbase publishes the developer documentation for implementing it on both the server and client side.

It is not the only proposal. Tempo, the payments chain incubated by Stripe and Paradigm, shipped a machine payments protocol alongside its mainnet launch in March 2026. Expect a standards fight rather than a coronation.

The numbers everyone quotes

Here is where the category gets interesting, and where most coverage stops reading.

Blockchain intelligence firm TRM Labs examined x402 activity from May 2025 onward across Base, Solana and Polygon. The analysis reported by PYMNTS found:

Two things fall out of that table immediately.

The average transaction is about 26 cents. Divide $52.7 million by 198.9 million transactions and you get the shape of the traffic: enormous numbers of tiny payments, which is exactly what the standard was designed for and exactly what no card network could process economically.

Less than half the value looked like commerce at all, and of what remained, the agent share was somewhere between one payment in a hundred and one in thirteen. That is the number to remember the next time you read that agents are already transacting at scale.

Why is the agent share a range rather than a number?

This is the most important methodological point in the whole category, and it is not a criticism of TRM's work. It is a limit on what chain analysis can do.

An AI agent deciding to buy a dataset, and a cron job that buys the same dataset every hour, produce identical records onchain. Same wallet, same signature, same call. There is no field in the transaction that says "a model chose this". Any estimate of agentic share therefore relies on heuristics about timing, variety and behaviour, and those heuristics have wide error bars by construction.

So when a vendor tells you that agentic commerce is growing at some precise percentage, ask how they separated agents from scripts. If the answer is onchain data alone, the precision is decorative.

What this is actually good for today

Strip out the agent narrative and a real product remains.

• Paid APIs without contracts. A data provider can price per call and serve anyone, with no sales cycle, no minimum and no invoice.
• Metered access to content and models. Inference, transcription, search results and archives can all be sold by the request rather than the month.
• Machine-to-machine settlement inside one company. Teams already use it to meter internal services without building a billing system.
• The groundwork for agent commerce, which is the bet everyone is actually making. The rail needs to exist before the traffic arrives, and it now does.

If you are building the agents that would eventually use this, our guides to autonomous AI agents and agent interoperability cover the layers above the payment, and our comparison of agent sandboxes covers where that code runs.

What to watch before you build on it

Refunds and disputes. A card payment can be reversed; a settled stablecoin payment cannot. If your agent buys the wrong thing, recovery is a commercial conversation, not a chargeback.

Spending limits. An agent with a funded wallet and a bug is a budget with no floor. Whatever you build, the limit belongs outside the agent, enforced by the wallet or the facilitator rather than by a prompt instruction the model can talk itself out of.

Security research is active. Academic work has already examined free-riding and data-leakage risks in x402 flows. Treat the standard as young, because it is.

Jurisdiction and reporting. Thousands of micro-payments a day are still payments. If your business needs to account for them, build that in from the start rather than reconstructing it from chain data later.

Conclusion

x402 is a genuinely good piece of engineering: a dormant status code, a stablecoin, one retry, and suddenly software can buy things. The standard works, the volume is real, and the major infrastructure companies are behind it.

The agent story attached to it is running well ahead of the evidence. Nearly 200 million transactions have settled, most of the value is not identifiable as commerce, and the portion that looks autonomous is a single-digit percentage at best. Both of those things can be true at once: the rail is being built now, and the traffic it was built for has not arrived yet.

If you sell data or inference, it is worth implementing today for the human-directed automation that is already paying. If you are waiting for agents to pay you, watch the share rather than the headline.

This explainer is an editorial synthesis of protocol documentation and published analysis read on 29 September 2026 and linked inline. The transaction and agentic-share figures come from TRM Labs research as reported by PYMNTS; we did not independently reproduce them. We did not implement or test x402 in production, and we hold no position in any asset mentioned. This is not investment or legal advice.