The line item reads $0.75 per 1,000 calls. Run the loop and it costs $1.50.
DeepSeek's new Search and Fetch endpoints are priced separately โ $0.75 per 1,000 calls each. A production agent does not search without fetching. It hits an index, receives ten URLs, then pulls the pages. That is two billable calls per retrieval cycle. Every headline I read last week quoted the $0.75 number. Every one of them was describing half a transaction.
Follow the gas, not the narrative. The gas here is $1.50 per 1,000 completed retrievals, and the gap between that number and the announced number is the first honest thing about this launch.
What actually shipped
Two standard endpoints, no architectural reveal. Search is an index query. Fetch is live retrieval. Alongside them sits V4.1-Flash-1017 โ the naming tells you what it is: a Flash-tier cost/latency checkpoint with an October 17 date stamp, not a generational architecture change. Continued training plus post-training optimization on the same base.
Billing terms matter more than the price. Failed requests, empty results, and timeouts are not charged. Five QPS per account, globally uniform pricing. Free until November 1, and the free window is scoped to three harnesses: DeepSeek Harness, Claude Code, and Codex. That list is not accidental. It is a customer-segmentation decision dressed as a launch promotion โ every name on it points at coding agents, the highest-frequency tool-calling workload currently in existence.
The asymmetry nobody priced
Search is cacheable. Fetch is not. Charging identical rates for both tells you which endpoint is subsidized.
Index lookups are read-heavy, batchable, and amortized against infrastructure that already exists. Live fetching means headless rendering, proxy rotation, anti-bot countermeasures, and per-request cost that scales linearly with volume. Same price, two completely different cost curves. That is not a pricing strategy โ it is a loss leader wearing one.
Crypto is the worst possible vertical to run that loss leader through, because crypto agent workloads are the most Fetch-heavy in the market. Nobody asks an agent to summarize a token. They ask it to verify a contract address against the explorer, pull the verified source, check the deployer's history, and cross-reference the governance forum before a transaction clears. That is four fetches and one search. The ratio is inverted.
I learned this the expensive way. In 2020 I built a Python script to sweep Uniswap V2 pools, and the compute was trivial โ the cost was in pulling token contracts and reading mint functions before the numbers meant anything. Fifteen percent of what I screened had hidden mint authority. The script didn't find that. The fetches did.
Where the cost actually moves
When a single retrieval drops from $0.005 to $0.00075, agent designers do not bank the savings. They change behavior. The optimization target shifts from "write the most precise query possible, cap the loop at three iterations" to "fire ten parallel queries, cross-validate, accept redundancy."
That behavior change is the real product, and it is worth more than the discount. Redundant retrieval is how you kill hallucination โ not with a better model, but with three sources that disagree in public. Volume goes up, revenue goes up, reliability goes up. Everyone wins except the incumbent selling one careful search for $7.
The 6.7-to-9.3x multiple you saw quoted is a comparison-set artifact. It benchmarks against Brave, Exa, Tavily, and Perplexity โ the agent-native newcomers โ and quietly omits Bing Search API, Google Custom Search, and SerpAPI, where cheaper tiers exist. It also omits the option nobody priced: self-hosted crawling. Any team already running indexers โ which in crypto is most of the serious ones โ faces a marginal fetch cost measured in queue depth, not dollars. For them, $0.75 is not cheap. It is redundant.
And the five QPS ceiling is the tell that $0.75 is an acquisition price. A single production agent averaging 10,000 calls a day runs at 0.12 QPS โ comfortable. A platform routing thousands of agents through one key does not fit, at any price. Scale customers will renegotiate. The published number is for the people who write the integration examples.
The part the launch post skipped
Two things sit outside the announcement, and both matter more than the price.
First, the exemption that supposedly de-risks this is a draft โ a proposed control list, not a finalized one. API and SaaS carve-outs are current text, not settled text. Anyone completing integration during the free window is building on a document that can still move.
Second, nobody has discussed the import side. A Western fund routing every retrieval query through an offshore endpoint is exporting intent data, and in crypto the intent is the position. Searching a specific contract address is not a neutral query โ it is a signal about what someone is about to buy. Model prompts can be scrubbed. Search queries are the intelligence itself.
I have spent this year building ETF inflow dashboards against on-chain exchange outflows, and the institutional lesson was simple: the data pipeline is the strategy. Handing the retrieval layer to a vendor whose cost structure you cannot audit, under a regulatory exemption that is not final, is a supply-chain decision disguised as a procurement one.
Watch the November 1 billing data. Not the press release โ the actual monthly spend of teams who integrated during the free window. That single number will tell you whether this was capacity monetization or penetration pricing, and those two things have opposite implications for everyone building on top.
One question worth sitting with: if a retrieval costs a tenth of a cent but returns a wrong contract address once in a thousand calls, what did you actually save?