State of x402 pricing in 2026: what agents actually pay

What the directory can measure

TOLL·402 stores provider, tool, network, minimum and maximum USD price, quote-check date, and verification state separately. That permits honest counts and price-band comparisons without pretending every service has the same billing unit. A web-search call, an image generation job, and a gateway request may all be priced per call while delivering very different amounts of work.

• Use minimum and maximum prices only when the endpoint or provider states them.
• Keep free tools distinct from tools whose price varies or is missing.
• Show the observation date beside any ecosystem-wide price claim.

How buyers should interpret a price

The current x402 buyer flow starts with an unpaid request. The resource server returns payment requirements, including scheme, network, asset and amount. That response is the moment to compare the charge against policy. Directory data helps choose candidates; the live requirement decides whether the agent should proceed.

A safer pricing policy

A buyer should combine a maximum amount per request, a maximum workflow budget, an allowed network and an allowlist of endpoints. The policy should stop rather than silently overpay when a price changes. TOLL·402 exposes price and network evidence so those controls can be configured before a wallet is connected.

What does an x402 price actually buy?

A price such as $0.001 per call looks precise, but the unit called a "call" can hide most of the economic difference between two services. One endpoint may return a complete extracted article. Another may return a search result ID that requires three more paid requests. A third may bill for a browser session whose value depends on whether the target page loads. The useful comparison is therefore cost per completed task, not the smallest number printed on a listing.

Before comparing services, write down the output the agent needs: ten cited search results, one normalized company record, a rendered page, or a generated image at a particular size. Then count every paid step required to reach that output. This turns a vague price comparison into a unit-cost comparison that can survive differences in product packaging.

• Identify the billing unit: request, result, token, page, job or session.
• Include follow-up calls needed to retrieve or poll for the final result.
• Check whether failed, timed-out or empty responses are charged.

What the July 2026 price sample shows

The editorial catalog is intentionally smaller than the full resource index because it contains listings with enough provider and product context to compare. Fifty-seven of 88 listings had a supported numeric minimum. One was explicitly free; 25 had a non-zero minimum no higher than $0.001; 18 were above $0.001 and no higher than $0.01; 10 were above one cent and no higher than ten cents; and three started above ten cents. The remaining 31 correctly stay unpriced rather than being treated as free.

This distribution supports a narrow conclusion: inexpensive entry points are common in the curated sample. It does not support the claim that most x402 work costs a fraction of a cent. The sample mixes gateways, single-purpose APIs, MCP tools, infrastructure and templates. Prices also come from different evidence dates and may apply to different operations within the same service.

A worked budget for an agent workflow

Suppose a research agent makes one search call, opens five results and retries one failed extraction. At $0.001 for search and $0.003 for each extraction, the happy path costs $0.016 and the retry raises it to $0.019. That is still inexpensive, but the per-call headline of $0.001 would have understated the workflow cost by nineteen times. If the agent repeats the workflow for 500 companies, the expected spend becomes $9.50 before any branching or human re-runs.

A useful budget has three levels. The request cap prevents a single surprising quote. The task cap limits one research job, including retries. The daily or wallet cap contains a bug that launches many jobs. The agent should calculate remaining budget before signing, not discover the limit after settlement.

• Per request: reject any requirement above the expected operation price.
• Per task: reserve enough for the planned call graph and a small retry allowance.
• Per period: stop all automated spending when the daily or wallet ceiling is reached.

Why the cheapest endpoint can cost more

Low price is valuable only when the response is usable. Poor extraction can force a second provider call, missing citations can require manual checking, and an unstable endpoint can consume the retry budget. Latency matters too: a service that is one tenth of a cent cheaper but regularly takes thirty seconds may be the expensive choice inside an interactive agent. These costs belong in the comparison even though they do not appear in the payment requirement.

For evaluation, use a small set of representative inputs and score completeness, structure, latency and failure behavior. Keep the test set constant across providers. A provider wins when it completes the job within the budget and quality threshold, not when it posts the lowest isolated quote.

How should an agent react when the quote changes?

The unpaid 402 response is a fresh offer, not an instruction to pay automatically. Compare its amount, asset, network, pay-to address and resource URL with the policy attached to that tool. A small price increase may still fit the task budget, but a new network or destination deserves a stop and review because it changes more than cost.

Do not silently increase a cap to keep a workflow moving. Record the old expectation, the new requirement and the decision. If the service supports usage-based pricing, authorize only the advertised maximum and retain the final settlement evidence so estimated and actual cost can be reconciled.

How the price distribution moved since July

The catalog roughly doubled between the first edition of this report and August. Both the count of listings exposing a supported minimum price and the count starting at a cent or less grew faster than the catalog itself, so the sub-cent floor is not an artifact of a small early sample.

The share of priced listings at a cent or less held between the two dates, which is the more interesting reading. Adding eighty-four listings did not pull the entry price up, and the cheapest band above zero grew in step. Entry pricing in this catalog is stable rather than drifting.

Where the corpus-wide benchmark sits

The editorial catalog is a curated layer, not the market. Across every discoverable priced route, the provider-weighted median is one cent per call, with a quartile range of $0.004 to $0.04 and a ninetieth percentile of $0.10. Capability medians range from half a cent for web scraping and inference up to a little over two cents for video and audio generation.

Those benchmarks collapse templated route families into distinct offers and let each provider contribute its own median exactly once, because a single bulk seller registering thousands of near-identical routes would otherwise set the number. Prices are what providers publish in their 402 quote. They are not evidence of demand, and a confirmed quote is not a settled payment.

• Provider-weighted median across all categories: $0.01 per call
• Quartile range: $0.004 to $0.04
• Cheapest capability medians: web scraping and LLM inference at $0.005
• Most expensive capability median: AI video and audio at $0.023

Related directory entries

Sources and methodology

TOLL·402 distinguishes public claims, registry discovery, unpaid quote checks and settled paid-call verification. Sources below support the visible claims; presence in a registry is not treated as verification.

1. TOLL·402 directory data — Curated listing and price fields used for the site.
2. x402 buyer quickstart — Authoritative current buyer flow and client packages.
3. x402 client/server concepts — Defines the live 402 requirements and retry flow.
4. The Going Rate for an AI Agent's API Call Is Two Cents — An independent provider-level and route-level analysis computed from this public dataset, including the July 10 median and live-quote pass rate.
5. x402 rate card — Provider-weighted per-call benchmarks by capability across every discoverable priced route.