Claude paid one cent to read a page: the experiment that previews how the web will charge AI agents

By Zendoric · September 17, 2026.

On September 15, five one-cent payments settled on a blockchain to unlock a web page. They were made by Suganthan Mohanadasan, an SEO consultant and co-founder of Snippet Digital, using his own AI agents and test money. One of those payments was executed by Claude Code, in the middle of a task, without him having to intervene.

The experiment, documented on his blog and picked up on Hacker News, consists of a site, paid.suganthan.com, that responds with HTTP code 402 "Payment Required" when an AI agent tries to read one of its pages. That code has existed in the HTTP standard since 1997, reserved "for future use"; it had never been used in practice. The x402 protocol, driven by Coinbase and now under the Linux Foundation umbrella, finally gives it a purpose: it turns the 402 into the first step of an automatic payment.

The mechanism has four steps: the agent requests the page; the server responds with the price (one cent in USDC, a cryptocurrency pegged to the value of the dollar); the agent signs a payment authorization and repeats the request; a "facilitator" —in this case, Coinbase's— verifies the signature, settles the transaction on the Base network, and the server delivers the content along with a public receipt bearing its own transaction hash. All without the agent opening an account or requesting an API key from the site's owner.

The most revealing part is not the payment itself, but the lock around it. Mohanadasan programmed his buyer agent with a cap of $0.05 per request and a daily allowance of $0.25; if the price requested exceeds those limits, the script refuses to pay and stops without ever signing anything. It is the same logic Cloudflare announced for its "Wallets", presented on August 4: a person puts up the money and sets the limit, an agent spends within it. For now those wallets only allow reserving a name; according to Cloudflare's own documentation, they cannot yet send, receive or hold funds.

A caveat the author himself stresses is in order: this is a demonstration. The USDC used are "testnet" tokens, with no real monetary value, and the payments go from his own agent to his own server. No real crawler from OpenAI, Anthropic, Perplexity or Google has yet paid to access someone else's page. An analysis by TRM Labs of 198.9 million x402 transactions, cited by Decrypt on September 13, further suggests that most of the protocol's current volume does not come from autonomous agents, but from scripted programs, load tests and transfers between accounts belonging to the same operator. It is a useful reminder, consistent with something we repeat at Zendoric: it is worth separating already-demonstrated capability from the infrastructure noise around it.

The experiment comes alongside a development with more real weight: as Search Engine Roundtable and Digiday reported in mid-September, Google confirmed a pilot program —the "AI contribution pilot"— that pays some publishers when their content "significantly contributes" to an answer generated in AI Overviews, AI Mode or the Gemini app. The monthly figure arrives in Search Console, but with no breakdown of how it was calculated; one publisher consulted by Digiday described it as "a black box". Therein lies the underlying tension: Google decides what the content is worth and how much it pays for it, whereas with the open protocol Mohanadasan is testing it is the publisher who sets the price and anyone can verify the payment on the blockchain without relying on the platform's word.

Between those two extremes several strategies are already under way. Large publishers negotiate direct licenses —News Corp, Financial Times, Reddit—. Smaller ones, with less bargaining power, often block AI crawlers with Cloudflare; according to Press Gazette (September 4), the blocking itself has in some cases served to bring AI companies to the negotiating table. Cloudflare, meanwhile, is pushing its own "pay per crawl" product toward a "pay per use" model, with pilots alongside Ceramic.ai and You.com, arguing that downloading a page does not prove its content was used in an answer. That is reasonable, but it shares the same problem as Google's pilot: the publisher can no longer check it for themselves.

Our reading is that this experiment, modest in practice, points to where the economic plumbing of the agentic web is heading. The implicit deal that sustained much of journalism and "free" content on the internet —the crawler indexes, the search engine cites, the user clicks and visits the page— breaks when the last link, the visit, stops happening: the agent reads, summarizes and answers without returning traffic. That leaves publishers, above all the small ones, trapped between two bad options: block, risking disappearance from the AI answers that today replace search, or give the content away in the hope that some platform will decide, unilaterally, to compensate them someday. That is the uncomfortable, short-term part of this transition, and it should not be glossed over.

The fact that x402 already has 40 members operating under the Linux Foundation since July 14 —among them Visa, Mastercard, Stripe, AWS, Google and Cloudflare itself— suggests this is not one consultant's technical curiosity, but the draft of a standard the payments and cloud industry is already willing to support. If that infrastructure matures and real crawlers —GPTBot, ClaudeBot, Googlebot— end up carrying their own wallets, what gets set in motion is not just "charging AI to read": it is a first layer of machine-to-machine micropayment economics, auditable and without a single platform setting the price for everyone, closer to the distributed abundance we defend at Zendoric than to the opacity of a payout decided behind closed doors. But that maturation has not arrived yet: today, the only AI paying in this system is the one the site's owner has programmed to do so.