DeepSeek launched its own web-search engine and two metered endpoints on October 17 — Search and Fetch, billed at $0.75 per 1,000 successful requests each, capped at five queries per second per account, with failures and timeouts going unbilled. Calls made from DeepSeek Harness, Claude Code and Codex are free until 00:00 Beijing time on November 1. Forkast notes the price sits six to nine times below established providers, all of which charge at least $5 per 1,000.

One day earlier, on October 16, China’s Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48 of 2026: a proposed export-control regime covering 23 AI technologies — models, agents, computing clusters — inside a broader list of 40 critical technologies spanning biotech and photovoltaics as well.

Two documents, twenty-four hours apart. One system.

The Exemption Is the Policy

The draft restricts overseas deployment of covered models. Stand up a restricted model on your own servers abroad and you will need a license. But access “provided through APIs, websites, apps and software-as-a-service platforms” is explicitly excluded, as is anything already released open-source or published academically.

Read that twice. Beijing is not restricting the export of Chinese AI capability. It is restricting the export of Chinese AI assets while leaving the rental market wide open — and then setting the rent.

This is not a loophole someone forgot to close. A government that can enumerate twenty-three technology categories down to the licensing tier is entirely capable of typing the words “API access” into a control list. The carve-out is the design. Weights and clusters are the things you cannot claw back once they sit on a disk in Texas. A metered endpoint is the thing you can throttle, reprice, or audit on a Tuesday afternoon from inside the jurisdiction, without anyone’s attorneys filing a single motion.

The export control protects the business model. It does not constrain it.

What Seventy-Five Cents Actually Buys

Compare the two numbers honestly. A Western retrieval vendor at $5 per 1,000 is selling you a service plus a margin. DeepSeek at $0.75 is selling you an index — its own, built in-house, which the company says spans hundreds of billions of pages and absorbs hundreds of millions of queries a day.

The per-call price is not the product. The routing is. Every agent that reaches for a citation through DeepSeek’s search becomes a downstream tenant of an index curated under somebody else’s rules, ranked by somebody else’s signals, refreshed on somebody else’s schedule. And the switching cost for a developer is 75 cents of invoice and an afternoon of SDK work — which is precisely why almost nobody will ever switch.

Cheap prices don’t create lock-in. They create inertia, which is cheaper to maintain and lasts considerably longer.

Note also the billing unit: successful requests only. It reads as generosity. It actually means the vendor is never paid to be reliable in the abstract, only reliable enough, and no uptime or latency commitment is published beyond the five-per-second ceiling. You are buying a best-effort dependency at a price point where nobody bothers to negotiate an SLA.

“We lost a renewal Tuesday because our per-call number had a comma in it,” a sales engineer at a US retrieval vendor wrote in a Slack DM between demos. “The customer didn’t benchmark anything. They looked at two columns and picked the shorter one.”

Who Has to Rethink Their Slide Deck

Two constituencies, and neither will enjoy it.

First, the decoupling caucus, which has spent three years insisting that export controls are the lever. The lever is real. They have been pulling the wrong end of it. If the durable chokepoint is the index rather than the weights, then the entire American argument over model export rules and chip restrictions is aimed at a layer already sliding toward commodity while the layer agents actually depend on — retrieval — gets priced into a foreign-controlled utility. The domestic vendors who would have to compete against $0.75 are the ones charging $5. Nobody in that debate has a slide for this, because the slide says controls are how you slow them down.

Second, the cheap-is-deflation crowd. The Hacker News thread on the pricing — 476 points, 232 comments — reads $0.75 as unambiguously good news about margins, and for anyone shipping an agent this month it is. It is also how a strategic dependency gets installed without a single legislator voting on it. Not through monopoly. Through a floor nobody can profitably undercut.

The stakes here are procurement, not geopolitics. If I’m right, the date to circle is not the licensing comment period. It’s November 1, when the free window for Claude Code and Codex calls closes. What happens on November 2 tells you whether $0.75 was a price or a hook.

Free through the end of the month. Then the meter starts — and a meter is a form of policy that never requires a signature.

Sources