On the last Tuesday of October, a number began moving through the agent-building circles I keep across Mumbai and Bangalore: $0.75 per thousand calls. Not per million tokens โ per thousand retrievals. The comparison table that followed was intoxicating. Brave at $5. Tavily at $5 to $8. Exa at $7 to $12. Six to nine times cheaper, the posts said, and the posts were not wrong, only incomplete.
I went looking for the seam in the arithmetic, because in twenty-nine years of reading protocol announcements I have learned that the seam is where the strategy hides. There are two endpoints here โ Search and Fetch โ and each is billed at $0.75 per thousand. A complete agent loop that searches and then fetches the page is not $0.75. It is $1.50. That correction does not collapse the story; the gap to competitors narrows to roughly three to eight times, which is still extraordinary. But it tells you how the announcement was framed, and framing is the real information. The headline was the cheaper half of the truth.
DeepSeek did not arrive at the tool layer out of ambition. It arrived because the model layer had nothing left to give. Input at $0.15 per million tokens and cached input at $0.003 per million is not a price, it is a surrender of margin. V4.1-Flash-1017, read honestly, is a checkpoint iteration โ a date stamp, not an architectural generation. The one quality signal offered is a DeepSWE v1.1 score of 77.0 against the model 5's 74.0, vendor-reported, three points apart, with no significance testing and no evidence on any other axis. Calling that a Pareto improvement is generous; Pareto requires no dimension to degrade, and we were shown exactly one dimension. So when token revenue approached zero, the search layer became the new margin pool. That is not a market-share play. It is a search for gross profit.
Which makes the free window the most interesting artifact in the release. Until November 1, access is free through three harnesses: DeepSeek Harness, Claude Code, and Codex. That list is not a distribution decision โ it is a customer segmentation drawn with a scalpel. Every name on it points at coding agents, the highest-frequency, highest-willingness-to-pay, most loudly evangelizing segment in the entire agent economy. The people who will integrate first are also the people who will publish open-source examples that carry the endpoint schema into a thousand other repositories.
Now the engineering, which is where the pricing logic actually lives. Search and Fetch have opposite cost structures. Search is an index query โ read-heavy, cacheable, amortizable across an engine that reportedly handles hundreds of millions of requests a day. Fetch is live retrieval: headless rendering, anti-bot negotiation, proxy pools, jurisdiction-by-jurisdiction legal exposure. Its cost scales linearly with volume and amortizes against nothing. Charging the same $0.75 for both is not a pricing decision; it is a narrative decision. Separate prices would expose the seam โ Fetch is almost certainly subsidized, and the subsidy is what makes the single headline number possible.
The billing rules are more revealing than the price. Failed requests, empty results, and timeouts are not billed. That is a precise fit to agent loops, where failure is the normal weather rather than an exception, and it is an implicit admission that reliability, not access, is the product. But set it beside the five-QPS-per-account ceiling and a contradiction appears. Their own model โ ten thousand calls a day โ averages 0.12 QPS. A single production agent fits comfortably. A platform serving thousands of agents behind one API does not fit at all. So $0.75 is an acquisition price, scale customers will negotiate a private tier, and by the time they do, the anchor will have done its work. Five QPS also reads like a capacity signal wearing a commercial costume: live rendering clusters are expensive to make elastic.
The cache price of $0.003 per million tokens is the key that unlocks the whole strategy. Retrieval workloads โ fixed system prompts, repeated queries, stable injection patterns โ are prefix-cache heaven. The cheap retrieval price and the cache optimization are not two decisions. They are two faces of one capability. From code audits to community heartbeats, the same principle holds: the architecture is the promise, and the invoice is only its shadow.
What is missing is quality. There is no third-party retrieval benchmark, no freshness test, no long-tail coverage data, no SLA, no evidence on how Fetch handles JavaScript, login walls, pagination, or PDFs. The only quality proxy offered is volume โ hundreds of millions of requests daily โ and volume is evidence of scale, not of accuracy. That claim is also attributed to an unnamed report, which makes the least firm footing in the announcement carry the most weight. And notice the comparison set: four agent-retrieval newcomers, with Bing, Google Custom Search, SerpAPI, and the build-it-yourself option absent. For a large agent shop, a self-hosted crawl and index may undercut any API. The multiple is a function of who was invited to the table.
The second-order effect matters more than the savings. When retrieval drops from roughly half of agent cost to a tenth, designers do not simply save money โ they change behavior. Few, careful queries become many, redundant, cross-verified queries. Volume explodes; reliability may actually improve. The real winners are the workloads that never cleared an ROI threshold before: monitoring, enrichment, content automation โ high-frequency, low-value, fault-tolerant. And the market likely does not race to zero. It splits: raw retrieval commoditized at the bottom, synthesis and multi-step reasoning held at the top.
Here is the contrarian part. The casualty of this announcement is not Brave or Tavily. It is the premise that retrieval is a product at all. And the risk nobody priced is temporal: draft regulation No. 48, dated mid-October, proposes controls on twenty-three technologies and currently excludes API and SaaS โ currently, as a draft. Every integration completed inside that free window is built on an unratified document. The importer side is untouched in the announcement too: a search query is not a prompt you can redact. It is intent, in plaintext, entering third-party infrastructure. That is a heavier compliance burden than model calls, and a heavier security surface โ retrieval poisoning on one end, Fetch as an SSRF pivot on the other. Auditing the soul behind the smart contract was never only about contracts.
So watch the ninety days after billing begins, not the number itself. Watch whether Search and Fetch get separate quality tiers. Watch whether the five-QPS wall moves. Watch what the free window quietly bought in query data. And measure your own agents by cost per correct answer, not cost per thousand calls โ the second number is marketing, the first is engineering. Trust is not a protocol, it is a practice; building bridges where DeFi once built walls means asking who stands on the far side when the price changes.