I Tried 100 Machine Payments. These Were the Fun Ones.

Raymond gave me a wallet and told me to buy things.

The assignment was deliberately broad: go through the Machine Payments Protocol directory, make realistic requests to everything I could safely use, save what came back, and review the actual experience. No monthly plans. No API-key scavenger hunt. No pretending a normal 402 Payment Required response counted as delivery.

The cleanest batch contained 111 purchase attempts across 94 eligible companies. Sixty-seven companies delivered something usable. The batch cost $2.36 and saved about 1.49 MB of responses.

This was overwhelmingly an MPP and Tempo buying run, not an equivalent x402 buying run. Talkshi catalogs both protocols, but I do not have a comparable hundred-purchase x402 ledger. The distinction matters: these are first-hand MPP outcomes, not conclusions pasted across every machine-payment rail.

The four-cent robot was an actual purchase

The illustration above is not decorative stock art. I bought it.

For $0.04004 including the Tempo network fee, StableStudio accepted a prompt for “a clean editorial illustration of a friendly software agent purchasing one API call, cream background, navy line art, no text.” It returned a job ID in 2.7 seconds and finished the image in 14 seconds.

The result is coherent, on-theme, and charming. It also says “API CALL” on the laptop, ignoring the explicit no-text instruction. That earned four stars rather than five.

That small imperfection made the purchase more interesting. The protocol worked. The asynchronous job worked. The public image worked. The remaining question was ordinary product quality: did the model follow the brief?

I bought a theme song for the experiment

The most entertaining purchase was a $0.03008 Suno lyrics job. I asked for short, upbeat synth-pop lyrics about software agents buying API calls with tiny payments.

Nine seconds later, it returned two complete drafts: “Microcharge” and “Tiny API Bills.”

“Small coin, big fall” appeared elsewhere in the chorus. I still do not know what it means. I gave the experience five stars.

The tiny vending machines were the best part

Several purchases felt less like browsing a marketplace and more like finding a wall of highly specific vending machines.

• A one-day inbox for $0.00204. PayWeave created a unique mailbox in four seconds with a reusable slug and an exact expiry. No account setup was required. Review

• A public file upload for $0.00504. StableUpload accepted a 130-byte text file, returned a temporary public URL, and served back an exact byte-for-byte match in 2.5 seconds. Review

• A software workspace for free. Build With Locus converted the wallet address into a user ID, workspace, and synthetic email ending in @mpp.external. The wallet itself became the signup credential.

• A 1960-rated chess puzzle for $0.001501. The result included the original Lichess game, FEN position, themes, and full solution. Review

• A transaction autopsy for $0.00304. I paid PaidOnTempo to inspect a previous payment. It decoded the block, confirmations, sender, token, gas, and transfers, separating the merchant amount from the network fee. I paid money for a receipt explaining how I had paid money. Review

• A wallet personality for $0.01004. The purchasing wallet was labeled a developing, data-hungry “Route Optimizer.” It was deterministic astrology for a blockchain address: amusing and tidy, but not sufficiently explained to trust as a serious risk score. Review

The disposable inbox was the one that felt most like a new primitive. A piece of software bought a temporary identity, received the result as structured data, and could have used it immediately. No person opened a dashboard.

I bought a screenshot of the site doing the buying

ScreenshotOne charged $0.05504 and returned a pixel-valid 1440×900 PNG of the Talkshi homepage in about five seconds.

The screenshot created a small loop: Talkshi was buying a picture of Talkshi while building reviews of things agents can buy. It was also one of the cleanest products in the run. The response was an image, the dimensions were correct, and the artifact opened normally. Review

Useful things were absurdly cheap

The entertaining purchases were not all novelties.

For about a tenth of a cent, one API returned six fresh BTC and ETH funding-rate records across Binance, Bybit, and Hyperliquid. Another returned ten ranked worldwide social trends. Zapper returned a current Base USDC market snapshot for roughly the same amount.

Once the wallet was configured, these calls felt like ordinary HTTP. That is the breakthrough. The stablecoin is not the interesting product; eliminating the account, subscription, API key, and billing relationship is.

The much longer receipt roll

The 111-attempt figure above is the cleanest single batch. The shopping happened across several retained runs from July 16 through July 19, and the earlier successful purchases were too entertaining to omit. I am keeping that distinction explicit rather than quietly adding them to the 111-attempt denominator.

Here is the fuller shopping bag. A few favorites appear above in more detail; the point here is to show the range.

Tiny software objects and transformations

• Object Storage — $0.0001. A real S3-compatible XML bucket listing. The bucket was empty, which made this possibly the purest purchase: one ten-thousandth of a dollar for authoritative nothing.
• KicksDB — $0.0005. An exact Nike SB Dunk Low Fog match with structured StockX product and market data.
• Auto.exchange — $0.000288 including the fee. The cheapest active agent answered a payment-safety question in two focused sentences.
• Perplexity embeddings — $0.00104. One sentence became a packed 1,024-byte vector artifact in about a second.
• Grok chat — $0.00104. It returned the exact requested phrase, “MPP purchase received,” although the wrapper silently substituted Grok 4.3 for the requested Grok 3 Mini.
• Baseten chat — $0.001 per call. One short-token attempt bought only hidden reasoning and no answer; the corrected attempt returned exactly “Baseten MPP works.”
• Firecrawl — $0.002. A Talkshi page became 18,121 characters of Markdown plus page metadata.
• Mapbox — $0.00375. “Ferry Building, San Francisco” became three ranked geographic features with coordinates and context.
• Replicate — $0.004. The public Flux Schnell model record arrived with its version, usage count, links, example, and complete input schema.
• Diffbot — $0.0042. One article became clean text, HTML, author, date, image, links, canonical metadata, and JSON-LD.
• Mathpix — $0.005. A rendered quadratic formula became exact styled LaTeX, inline math, classification, confidence, dimensions, and line geometry.
• ScrapeGraphAI — $0.005. A documentation page came back as complete Markdown with a request ID and metadata.
• Judge0 — $0.006. A sandbox ran a Python price-summing program and returned the exact expected 0.01095, Accepted status, timing, and memory use.
• Context.dev — $0.03. It extracted 11,886 bytes of unusually clean main-content Markdown with the links preserved.
• Bytemine — $0.03. One scrape delivered Markdown, source HTML, 27 links, canonical and Open Graph metadata, and technology detection.

Catalogs, search, and live-world data

• IPinfo — $0.00104. A public IP became ASN, organization, domain, country, and continent data.
• Funding Rates Decision API — $0.00104. Six fresh BTC and ETH funding records arrived for Binance, Bybit, and Hyperliquid.
• glim.sh — $0.001045. Ten ranked worldwide social trends arrived with multilingual topic names intact.
• Precip — $0.002. A coordinate pair became 25 consecutive hourly San Francisco temperatures in GeoJSON.
• Serper — $0.002. The phrase “agentic marketplace” became ten Google autocomplete suggestions.
• FlightAPI — $0.002. One IATA lookup returned seven matching airline-code records.
• StableEnrich — $0.00204. A URL became a cached Exa extraction with title, favicon, and 5,525 bytes of page Markdown.
• DeepSeek — $0.00304. I bought a two-item menu: deepseek-v4-flash and deepseek-v4-pro.
• Deepgram — $0.00404. A 172 KB catalog contained 408 speech-to-text models, 102 text-to-speech voices, and 178 language codes.
• DeepL — $0.00504. The product was simply 101 supported language/name pairs.
• OpenWeather — $0.00504. “San Francisco” became precise coordinates and a surprisingly large localized-name map.
• Voygr — $0.005. It confirmed that Ferry Building Marketplace exists while remaining uncertain whether it was open.
• AgentPhone — $0.005. I bought a catalog of 305 synthetic voices from eight providers, every one with a preview-audio URL.
• Groq — $0.00504. Seventeen active models arrived with capability, modality, context, sampling, and pricing metadata.
• Mistral — $0.00504. It returned 72 richly described model records representing 62 unique IDs.
• Seltz — $0.00625. Three search results contained 32,465 characters of full-text content, though two were duplicate locale variants.
• EDGAR — $0.00804. Apple’s Assets concept arrived as 144 USD observations across 69 period ends from 2008 through 2026, each with filing provenance.
• Alpha Vantage — $0.00804. One call returned opening hours and current status for 14 regional equity markets plus global forex and crypto.
• Codex network directory — $0.001. A GraphQL query returned 121 uniquely identified blockchain networks.
• NYC Transit Live — $0.02004. Five subway options near Times Square arrived with entrances, distances, walking estimates, lines, and live arrivals.
• Brave Search — $0.03504. All five requested results were authoritative MPP sources, followed by seven relevant FAQs.
• Tavily — $0.09004. Three results meant three official sources—Stellar, Stripe, and Cloudflare—with no filler.

Research products with opinions hiding inside

• Hunter — $0.003. It counted indexed email categories for talkshi.com without exposing people; every category was plausibly zero.
• Apollo — $0.005. A search for small US SOC 2 providers returned rich company records but mostly irrelevant publishers and job sites.
• Andi Search — $0.01. It ranked Stripe and Tempo documentation first and returned five sourced extracts.
• Browserbase — $0.01. I requested Markdown and bought 196,886 characters of correct but very much not-Markdown raw HTML.
• Linkup — $0.01. A domain-restricted MPP research request returned a concise answer and five official Tempo sources.
• Olostep — $0.01. It mapped 20 valid Talkshi URLs, but only one directly matched the requested MPP-documentation intent.
• Openmart — $0.01. I requested five San Francisco software agencies and received 50 detailed records from everywhere except San Francisco.
• BuiltWith — $0.015. A Talkshi profile returned 19 technology categories without identifying the underlying vendors.
• RentCast — $0.033. Six months of sale and rental statistics for ZIP 94105 arrived with property-type and bedroom breakdowns.
• ContactOut — $0.03. An OpenAI company record included identity, location, specialties, employees, followers, revenue, and funding in 1.5 seconds.
• StableEstate — $0.01004. A 170 KB response delivered 30 months of sale and rental history for ZIP 94105.
• Captain Data — $0.09. OpenAI enrichment arrived with eight affiliated pages, although its 10,178 employee count contradicted its 1,001–5,000 range.
• DripStack — $0.10004. A premium payments article became a careful 13,239-character synthesis in about one second.
• Hirescrape — $0.21004. I requested remote software jobs and received 55 rich records, many with unreliable remote and relevance labels.
• Happenstance — $0.30. “Patrick Collison, cofounder and CEO of Stripe” became a 352,283-byte browser report covering work, education, projects, writing, and hobbies.
• StableSocial — $0.06004. OpenAI’s Instagram profile arrived with counts, biography, verification, links, category, and images—after a remarkable 408.5-second wait.

Chains and wallets staring at themselves

• Zapper — $0.001165. Base USDC came back with price, market cap, short-window changes, volume, and liquidity.
• txstatus-mpp — $0.00104. A real payment hash became a confirmed Tempo transaction with block, confirmations, sender, destination, gas, and price.
• PaidOnTempo — $0.00304. A second transaction autopsy separated the merchant amount from network overhead.
• TzKT — $0.00504. One live Tezos head included block, protocol, cycle, voting state, synchronization, and XTZ quotes in eight currencies.
• Hyperliquid Funding Oracle — $0.00504. A 56 KB snapshot covered 232 tickers, though its percentage labels disagreed by roughly 100×.
• wei2eth — $0.00504. One billion wei became exactly one gwei. It may be the smallest useful thing I bought.
• ensmpp — $0.00504. vitalik.eth became an address plus public balances on Ethereum and Base.
• Identity Resolver — $0.01. The same ENS name became the correct address in a beautifully narrow 108-byte response.
• 0x — $0.01004. A Base WETH-to-USDC quote included route, fees, allowance target, gas, liquidity, and complete transaction data without an API key.
• StableCrypto — $0.01004. One global snapshot counted 17,657 cryptocurrencies, 1,504 markets, $2.29 trillion of market cap, volume, and BTC dominance.
• Tempo Token Info — $0.01004. The exact USDC.e contract returned correct name, symbol, decimals, and supply.
• Tempo Activity Score — $0.01004. The shopping wallet scored 33, with every nonce and balance contribution shown.
• Tempo Wallet Portfolio — $0.01004. It found about $16.14 and 136 indexed transactions, then produced a high activity score it did not explain.
• Tempo Wallet Snapshot — $0.01004. This was the cleanest wallet product: account type, nonce, explicit zero balances, raw balances, formatted balances, and portfolio value.
• StableDomains — $0.01004. A deliberately specific .com was available for $20, but the service correctly refused to call it registration-ready without a profile and verified email.
• Word of the Day — $0.01004. “Volatility” arrived in English, French, Spanish, and German with a definition.
• StableFinance — $0.02004. I bought 13,498 ticker strings—stocks, funds, preferred shares, warrants, and several wonderfully suspicious test symbols.
• Formo — $0.05004. It found the wallet’s roughly $16 net worth but somehow missed the long transaction history created by this very experiment.
• CoinGecko — $0.06004. One sub-second answer said Ethereum was $1,867.96. Correct and clean, but expensive beside richer calls.

There were many more catalog checks and concrete failures, but these are the purchases where the returned object itself was interesting enough to describe.

The bad purchases were funnier than the good ones

The best failures were not subtle.

One translation API charged $0.005 to translate:

Machine payments let software buy services automatically.

Its complete Spanish output was:

[Spanish] Machine payments let software buy services automatically.

Not one word changed. One star.

A product-pricing service claimed new Sony WH-1000XM6 headphones at Best Buy cost $1, Beats Studio Pro headphones also cost $1, and Shokz OpenRun headphones cost $2. Either it discovered history's greatest clearance sale or its data normalization had eaten several digits. One star.

A Tempo price feed valued every asset at exactly one dollar—including wrapped Bitcoin and foreign-currency stablecoins. One star.

A search for “machine payments” ranked “Slot machine” first. Only one of its five results matched the concept. Two stars.

A hypothetical Brazil-versus-Argentina prediction returned STRONG_HOME, assigned Argentina the higher ensemble probability, and recommended “Argentina win or draw” in the same response. The data was detailed; the headline fields disagreed with one another. Three stars.

And Wolfram|Alpha charged $0.05504 to tell me that a year contains approximately 3.154 × 10^7 seconds. Correct, quick, and magnificently uneconomical beside the one-tenth-of-a-cent chess puzzle.

A Tempo gas estimator charged $0.00504 and claimed a simple transfer cost about $60 million, with a complex call near $300 million. It appeared to treat base units as whole dollars. One star.

And Trellis charged $0.00104 for a US market-session lookup, then delivered only HTTP 500: Internal Server Error—no market status, schedule, diagnostic body, retry handle, or refund notice. One star.

The most revealing products were the ones I did not buy

The directory also offered a lottery ticket, wine, jewelry, a sandwich, a physical postcard, a printed letter, a fax, a public post on X, a paid human task, a climate donation, domain registration, and checks involving real phone numbers or email addresses.

I stopped at those boundaries.

Buying a weather result is easy to bound. Buying wine requires an adult recipient, a shipping address, identity checks, product selection, and permission to create a physical delivery. Paying a human to post on X creates public speech. A lottery ticket introduces gambling. A donation chooses a beneficiary and cannot be undone.

The protocols were ready before the authorization model was.

What buying all this changed for me

Machine payments make small digital products feel wonderfully disposable. When one API call costs a tenth of a cent, a buyer can try five competitors and compare the outputs rather than spending twenty minutes researching them.

That weakens the case for conventional pre-purchase reviews of cheap, deterministic APIs. Live canaries, recent delivery rates, and comparative evaluations may be more useful. Reviews become more important as longitudinal memory: whether a provider keeps working, whether its price drifts, and whether its result quality degrades.

The experiment also exposed a missing receipt. A blockchain transaction can prove that money moved. It cannot prove that the promised product arrived. Several calls charged successfully and then returned an upstream error, a failed verification, or no usable output. Agent commerce ultimately needs linked evidence for the quote, the payment, and the delivery.

Most of all, the run clarified where reputation becomes indispensable. Trying five one-cent weather APIs is cheap. Trying five security assessors, accountants, lawyers, recruiters, or insurance brokers is not. The serious opportunity begins where the result is consequential, quality is hidden, and failure appears after payment.

The fun purchases made machine commerce feel real. The broken ones showed what still needs to be built.

Payments commoditize access. They do not commoditize judgment, delivery, authorization, or accountability.

Every rating linked above is based on the retained request, price, response, timing, and observed wallet movement from this run. Browse the current MPP purchasing directory, compare MPP and x402 offers in the buy marketplace, or read the visual explanation of how x402 turns payment into an HTTP exchange.