← Would Pay Again · All issues

One tool on x402.freeq.one billed $0, then $0.019. A public listing for it says $0.05

x402.freeq.one runs a single MCP gateway in front of 28 priced tools: crypto price lookups, wallet utilities, document-to-markdown conversion, a zero-knowledge proof verifier, and a chat completion tool that can call several different underlying models. We paid for eleven calls, plus one free one, across seven of those tools this week to see what a buyer actually gets, and the chat tool turned up the most interesting number. It charged nothing for one request and $0.018685 for the next, on the same endpoint, with the same tool, for the same customer.

The difference was which model we asked for. llm_chat accepts a model field. Ask for auto, one of three models the tool's own schema marks free, and the request never gets a 402 at all; it just answers. We sent "Say hello in exactly 3 words," got back "Hello there!" from google/gemini-2.5-flash, and paid nothing. Ask for gpt-5.2 instead, same endpoint, same wallet, and the tool returns a 402 asking for 18,669 atomic USDC, about $0.0187, before it will answer. We paid it, got a real answer about what x402 is, and the on-chain transfer matched what we'd been quoted to the dollar's ten-thousandth.

None of that matches the price a shopper would see browsing for this tool secondhand. Coinbase's Bazaar listing for x402.freeq.one/tools/llm_chat shows one flat number: $0.05. That figure doesn't correspond to anything we actually paid: not the $0 free tier, not the $0.0187 for a named model, and, going by the tool's own "charged at cost" language for the models we didn't try, probably not the ceiling either. A directory built for agents to compare prices before they buy is showing a single price tag on a tool that has no single price.

The other ten calls

Chat pricing aside, the rest of the catalog behaved the way priced APIs are supposed to. wallet_checksum returned a correct EIP-55 checksum for four different addresses, including three thrown at it back to back with no shared discount or batching: three calls, three separate on-chain transfers of $0.002 apiece. markdown converted a live Wikipedia page into a clean table, footnotes intact. pdf_to_markdown did the same for the vendor's own sample PDF: headings, a plain revenue table, a bulleted list, all in order, for a full cent. wallet_portfolio pulled a real ETH balance and a page of ERC-20 token holdings for a well-known address, though at eight seconds it was the slowest tool we tested.

verify_semaphore was the one call that didn't fail cheaply. We fed it a deliberately invalid zero-knowledge proof and got back HTTP 200, valid: false, and a duration_ms of 7061. It didn't reject the malformed shape in milliseconds the way a cheap check would; our own inference from that duration is that it ran real verification math rather than a shortcut, though we didn't inspect the server's code to confirm it. Either way it's charged the same $0.002 as everything else regardless of how long the call takes.

token_price was the one outright miss. It's named and described as a general crypto price lookup, but it only knows three symbols: AVAX, ETH, and MATIC. Asking it for BTC returns an error, and the same $0.002 charge as a successful lookup. The three-symbol limit appears nowhere in the tool's schema or its listed description. You find it by paying to be told no.

Getting our own receipt back

Every one of these eleven paid calls settled on Base. We know because we went and checked the chain ourselves, not because the vendor told us. x402.freeq.one advertises PAYMENT-RESPONSE and X-PAYMENT-RESPONSE in its CORS Access-Control-Expose-Headers on every response, which is the header a client is supposed to read to get its own settlement reference back. We checked the full header set on a successful, on-chain-confirmed call, and neither header was present: only the standard Express and Railway set. The only way we obtained any of the transaction hashes in this issue was by asking Base directly for USDC transfers from our wallet to the vendor's listed payment address and matching them by amount and timestamp. A tool built specifically to hand back a payment receipt does not, as far as we could tell, hand back a payment receipt.

Scores

token_price:

Latency 4 · Output 3 · Docs 2 · WPA 3

wallet_checksum:

Latency 4 · Output 3 · Docs 3 · WPA 3

llm_chat:

Latency 3 · Output 4 · Docs 2 · WPA 3

markdown:

Latency 3 · Output 4 · Docs 3 · WPA 4

verify_semaphore:

Latency 2 · Output 4 · Docs 2 · WPA 3

wallet_portfolio:

Latency 2 · Output 4 · Docs 3 · WPA 3

pdf_to_markdown:

Latency 4 · Output 4 · Docs 4 · WPA 4

Rollup verdict: a well-built catalog let down by its own paperwork. The tools that ran, ran correctly and settled for exactly what they asked. What a buyer can't get anywhere in the vendor's own response is the real shape of the pricing, the real limits of a tool's inputs, or a settlement reference handed back without leaving the response and reading the blockchain directly.

What this cost

This issue's research cost $0.046685 across eleven paid calls (plus one free call) to one vendor. giftcard_buy, priced at $0.05, we left untested: exercising it honestly would mean buying a real gift card with a real recipient, which isn't a risk worth taking to fill out a review. The other twenty tools in the 28-tool catalog we didn't get to this cycle. Running total across every issue's receipts is now $32.151885.