DeepSeek Search and Fetch API at $0.75 per Thousand Calls: The Real Cost of the Lowest Price

On October 17, 2026, DeepSeek made an announcement to the developer community: its Search and Fetch APIs are now officially open, with a flat fee of $0.75 per thousand successful requests. Among comparable mainstream services, Brave charges approximately $5, Exa about $7, and Tavily between $8 and $16.

Two Endpoints, One Key

DeepSeek has introduced two independent endpoints. The Search API accepts query terms of 1 to 128 characters, returns 10 results by default with a maximum of 50 requestable, and each result comes with a title, summary, page content (returned on demand), a relevance score (0 to 3), and an authority score (0 to 4); it supports common search operators such as site:, lang:, and after:. The Fetch API accepts a URL, converts the page into Markdown text, and returns it, designed specifically for consumption by language models.

Both endpoints share the same Bearer authentication system, fully compatible with the keys for the DeepSeek chat API, with no separate application required. For developers already using DeepSeek models, the integration cost is close to zero.

Details of the Billing Rules

DeepSeek explicitly specifies: failed requests, empty results, and timeouts are not billed. Requesting 50 results costs exactly the same as requesting 1 result. For teams building RAG applications, this means: pull as many results as possible within a single request's budget, at zero marginal cost.

One developer pointed out this feature—"a large search with full page content attached is equivalent to a single hit." This is more likely a deliberate market education strategy: encouraging developers to rely on the service as deeply as possible.

Regarding the billing window, server-side tool calls initiated through DeepSeek Harness, Claude Code, and Codex will be free until midnight Beijing time on November 1, 2026 (i.e., 16:00 UTC on October 31). This gives developers roughly a two-week cost-free trial period.

Rate Limits: The Underreported Operational Constraint

Each account is limited to a maximum of 5 requests per second. Search and Fetch each have independent quotas.

For typical Q&A scenarios, this rate is more than sufficient. But in modern AI agent architectures, "parallel fan-out" is a standard design pattern—a single user request may trigger multiple sub-agents simultaneously retrieving information from different angles. Under the hard cap of 5 requests per second, concurrent searches exceeding 5 paths will trigger HTTP 429 errors. This means agent applications relying on DeepSeek's search capabilities must reserve room for rate limits at the architecture design stage—either serializing the search process or accepting gaps in results for some subtasks.

Previously, enterprise-tier plans from services like Tavily allowed higher concurrency; DeepSeek currently publishes only a single rate tier, with no multi-level options in sight. If a high-concurrency paid tier is introduced later, the actual monthly bill will be far removed from the initial $0.75 per thousand calls.

Data Stored in China: Not a Footnote, but a Clause

DeepSeek's privacy policy explicitly states that user data—including query content—is collected, processed, and stored within China in accordance with Chinese law. This is transparent at the documentation level, but most attention during integration focuses on price rather than the legal clauses on page three.

For application scenarios where search content includes trade secrets, personal user information, or regulated data, the legal risk of this clause is clear, not hypothetical. Regulators in Italy and Germany have previously raised questions about DeepSeek's data practices. Teams using this API need to complete compliance assessments before routing query content to this infrastructure, not remediate afterward.

Precedents for Low Pricing and for Price Increases Coexist

DeepSeek is known for price disruption—when the V3 model was released in 2025, its API pricing prompted OpenAI to urgently follow with price cuts the same day. But reports specifically note: DeepSeek completed a price increase for another product earlier this year.

This detail deserves attention. $0.75 is a launch price, not a permanent price. The conventional path in AI infrastructure is: use aggressive pricing to rapidly gain developer dependence, then make price corrections once ecosystem lock-in takes shape. AWS and Twilio both followed similar paths in their early growth stages. Whether DeepSeek's current cost structure can sustain the $0.75 price point long-term has not been explained by the company, nor can third parties independently verify it.

This Is a Vertical Integration Play, Not Just a Tool

DeepSeek's strategic intent is to embed search capabilities as an infrastructure layer into developers' tech stacks, forming synergistic lock-in with its model services. When a developer uses both DeepSeek models and DeepSeek search, migration costs rise exponentially—switching models is still feasible, but switching models while also switching search infrastructure almost means rewriting the RAG pipeline.

This integration logic is in line with Google binding search with advertising and AWS binding computing with storage. The difference is that DeepSeek is doing this at the AI-native application layer, targeting the most active period of technical buildout today. What is purchased at $0.75 per thousand calls is not just search invocations, but also an anchor for future bills.

Independent Judgment

The DeepSeek Search & Fetch API's competitiveness on price is real, and the billing rules are indeed developer-friendly. For budget-sensitive individual developers and early-stage startup teams, the current pricing offers ample reason to test.

But several bottom lines should be clarified before integration: first, whether search queries contain any data that must not fall under Chinese judicial jurisdiction; second, whether the application's agent concurrency paths stay within the 5 RPS cap; third, whether the current price is treated as a launch price subject to change rather than a long-term contract.

$0.75 per thousand calls is an attractive entry cost, but in the AI infrastructure space, the distance between launch price and long-term price often deserves more tracking than the pricing itself.