DeepSeek Made Search Almost Free. The Number It Returns Is the Bigger Story

In Claude Code, you switch the backend model to DeepSeek and ask it to check the latest version of a library while writing code. It issues a search, gets back a few results, reads one, and keeps writing. The action is so light that developers barely notice it. But coding agents depend on search far more heavily than chat scenarios do. Look up one API parameter incorrectly, and the next dozens of tool calls, code generations, and test fixes are built on an outdated premise. One bad search does not just add one useless piece of information. The entire task chain has to start over.

DeepSeek has opened its self-developed search API, with both Search and Fetch priced at ¥5 per 1,000 successful requests. To understand this price, you cannot look only at the number “five yuan per thousand calls.” You need to see where it places search, what billing basis it uses, and what filtering tools it leaves to downstream developers.

Pricing: a tenfold unit-price gap, and another layer in billing basis

At ¥5 per 1,000 requests, the per-call price is ¥0.005. Bocha, which the project at hand is using, has a list price of ¥0.036 per call, or ¥36 per 1,000. Tavily’s pay-as-you-go price is $0.008 per call, roughly $8 per 1,000, or more than fifty yuan per thousand calls. Brave is about $5 per 1,000, and Exa’s standard tier is $7 per 1,000. On unit price alone, DeepSeek is more than seven times cheaper than Bocha and about ten times cheaper than the Tavily family.

The more important difference is the billing basis. DeepSeek charges only for requests that return HTTP 200 with non-empty results. No results, timeouts, and failures are not billed. During debugging, large numbers of empty queries do not cost money for nothing. Search and Fetch can also be completed in the same request. When page content is needed, it is returned directly, instead of billing one search and then multiple fetches. In a traditional RAG pipeline, a task might require 10 searches and 10 fetches, creating 20 billable events. Under this design, it might be only 10 successful searches, with fetching already included.

In practice, the cost gap is another layer larger than the unit-price comparison. For Agent projects that frequently go online, this gap is enough to change architectural choices. The trade-off that used to be made between search quality and call cost is now first pulled by cost into a new feasible region, and quality is discussed afterward.

But ¥5 per 1,000 requests has a clear boundary: 5 QPS per account, with Search and Fetch counted separately. At 5 QPS, the absolute maximum is about 430,000 requests per day. That is enough for individual developers and early-stage Agent projects. An application doing concurrent retrieval for real end users will hit this ceiling early in its growth. This limit shows that the current target users are individual developers and small teams, not a replacement for enterprise search procurement. Price opens the door; QPS draws the frame very clearly.

The API design points to the same thing. The summary field defaults to 600 to 800 characters. The documentation’s exact wording is “suitable for model reading.” snippet is 100 to 200 characters, “suitable for front-end display.” Two sets of summaries are provided separately, and the default is clearly for model use. It supports syntax such as site:, lang:, after:, and before:, and there is an ori_question field that can pass in the user’s original question to prevent the query rewritten by the Agent from losing the user’s original intent. Fetch returns Markdown directly, saving a layer of HTML cleaning. Anyone who has built RAG knows how annoying this step is. Put together, these details point not to a general search portal, but to a retrieval front layer for Agents.

authority: the convenience of one if line, and the gate of a single integer

Each Search result returns two scores. rerank_score is from 0 to 3, combining relevance, authority, and timeliness. The documentation states that 1.5 or above counts as a high-quality result. authority is an integer from 0 to 4, with -1 meaning unknown. A score of 3 or above counts as a “highly authoritative result.”

Traditional search engines internally have site quality scores and page quality scores, but these scores have always been a black box. Baidu does not tell a querying user how authoritative a page is, and neither does Google. The SEO industry guesses, or infers from ranking changes. DeepSeek writes the score into the API response. A developer can add one line, if authority >= 3, in an Agent to filter, and no longer needs to maintain a “trusted-site whitelist.” A whitelist is slow to update, narrow in coverage, and quickly outdated. A public discrete score lets downstream developers perform batch filtering at extremely low cost. This is a concrete engineering benefit.

The other side of convenience is a gate. In traditional search, a personal technical blog with a low authority score but solid content could at least appear on the second page and be found. Human users scroll down, sense that something is off and switch keywords, and click into a lower-ranked page to judge for themselves. Agents do not do these things. An Agent program does not “sense that something is off.” If large numbers of Agents decide what to read and what to discard based on authority >= 3, this integer decides more than ranking. It decides whether a web page is eligible to enter the model context. Under threshold filtering, that personal technical blog has no second page. It simply does not exist.

Authority judgments naturally favor large sites, old sites, and institutional sites, because the authority proxy metrics that can be computed at scale, such as domain history, backlink structure, and institutional endorsement, are systematically missing for small sites and personal blogs. A great deal of first-hand technical experience exists precisely in personal blogs, forum posts, and GitHub issues. DeepSeek may not intend to exclude them, but this exclusion will appear as a byproduct of how the score is constructed. The documentation does not explain the dimensions used to calculate authority, nor does it provide an appeal or correction channel. Not disclosing the algorithm is understandable commercially. But when the score is used by downstream developers as a filter, a non-appealable integer is in fact exercising discretionary power that has not been publicly defended.

A 63-question test, which the original draft did not link, found that among 21 results manually labeled as spam or gray-market content, none of the authority scores returned by the API exceeded 2, and 13 were at the lowest level, 1. Official institutions and academic encyclopedias averaged around 3.4. This shows that the score discriminates against traditional black-hat SEO and low-quality content. The same test also found that, when facing GEO content published through high-authority sites and written for LLM-facing features, the existing filtering and scoring mechanism “basically cannot distinguish” it. authority is good at identifying content with poor provenance. It is not good at identifying content with good provenance but impure motives. The latter is precisely the main placement method of the GEO gray market today. Instead of building their own spam sites, operators blanket high-authority platforms with “review-like” articles.

Anti-GEO: data scale is an advantage; the capability boundary needs to be stated clearly

GEO has become wildly popular in China over the past two years. People in content marketing have switched to taking GEO contracts. The business is to mass-produce “review-like” articles, blanket various platforms, and make LLMs mention the client’s brand when answering “which brand in such-and-such category is best.” This is especially common in consumer questions. Click into the sources cited in the answers and you find a dozen or so articles spun from the same template.

DeepSeek writes filtering for SEO cheating, GEO manipulation, and low-quality AIGC into its announcement, and pairs it with a public authority score. The meaning is clear. What it wants to do is not simply sell search calls. It wants to become the trust layer when AI applications obtain external information. This positioning has evidence behind it. DeepSeek’s web endpoint and API handle hundreds of millions of search requests every day. At this scale, it can see earlier than small vendors which content is being placed in bulk and which sites suddenly sprout large numbers of similar pages. Identifying GEO is essentially a data-scale problem. The more it sees, the easier it is to recognize patterns. This is harder for small search API vendors to catch up with than price.

But “what it can identify” needs to be stated precisely. From existing tests, the filtering mechanism has shown verifiable effectiveness against traditional SEO cheating and low-quality AIGC. Against GEO content on high-authority sites, its discriminating power is clearly insufficient. This is not a problem that “seeing more” will automatically solve, because the placement strategy of GEO content itself exploits a blind spot in authority judgment. It piggybacks on platforms whose authority is naturally high. If mass-produced “review” articles appear on a site whose own authority is 3, the difficulty the search layer faces is not “failing to recognize that this is bulk content.” The difficulty is “recognizing the bulk pattern, yet having the host authority signal override the bulk signal.”

GEO defense and authority scoring are not two independent functions here. They interfere with each other. DeepSeek placing “anti-GEO” and “public authority” side by side in the announcement is right in direction, but not symmetric in capability. Its defense on the traditional SEO battlefield is credible. Its defense on the high-authority GEO battlefield has not yet been proven. The latter is precisely the part that most affects the credibility of model output in consumer queries and business-decision queries. authority is more like a provenance signal. It can tell you where the page comes from. It cannot fully tell you why the page exists.

Platform position: search moves from an add-on tool to a server-side precondition for the model

There is another easily overlooked detail in the announcement. When users use the DeepSeek API in DeepSeek Harness, Claude Code, and Codex, they are already accessing the Search and Fetch interfaces through server-side tool calls. Currently there is no charge. Starting November 1, billing will be per call.

This shows that the search capability has been running online for a while, and running inside other people’s coding tools. When a developer switches the backend model to DeepSeek in Claude Code, the moment it goes online to check documentation, it is using DeepSeek’s own search. The tool still belongs to someone else, but the data flow already belongs to DeepSeek. Coding agents depend on search far more heavily than chat scenarios. Checking the latest version of a library, a parameter change in some API, an issue for some error: these needs have strong timeliness and strong accuracy requirements. One bad search and the next dozens of tool calls are basically wasted. Search quality is a hard requirement for coding agents, and DeepSeek is first getting search working in this scenario. The path is well chosen.

More long-term, when search moves from an “add-on tool” to a “default precondition on the model server side,” the commercial value of the retrieval layer is no longer just selling call counts. It becomes the unavoidable path through which Agents obtain external information. The role of ¥5 per 1,000 requests is to make this path economically almost without reason for replacement. Once the retrieval habits of large numbers of Agents are locked into this interface, downstream use of the authority score will gradually move from “optional filter” to “default practice.” By then, the window for discussing the fairness of the score will be much narrower than it is now.

Conclusion: treat authority as a hint, not a verdict

DeepSeek’s opening of the search API this time has a price shock. But understanding it only as “the price butcher is back” would miss the more important change. It pushes search one step from a “ranking service” toward an “admission service for Agents.” It not only decides how results are ranked. Through a public discrete score, it also provides downstream developers with a low-cost filtering method that can run in batches.

This design is clever in engineering and clear in business. But it elevates authority judgment from an internal ranking factor of a search engine into a hard threshold that external Agents can directly execute. Regarding how this threshold is set, how to appeal, and whether it is fair to different content types, public information is far from sufficient. Once a score with no public standard is treated by large numbers of developers as a default filtering condition, it will in practice decide the range of information read by huge numbers of AI applications.

For developers, the more prudent approach is not to directly adopt authority >= 3 as an all-or-nothing threshold. Treat it as one of several signals. For sources with low authority, do not automatically discard them. Require additional cross-validation: consistent information from multiple independent low-authority sources, or alignment with factual fields from high-authority sources. This way, you can still use DeepSeek’s ability to suppress spam without completely handing the visibility of a small technical blog to an integer produced by an undisclosed algorithm.

Treat authority as a hint, not a verdict. Before DeepSeek provides more detailed score explanations and feedback mechanisms, this may be the most pragmatic strategy for balancing cost and information diversity. Competition among search APIs is moving from “who is cheaper and faster” to “who can provide a credible external information entrance for Agents.” Price decides whether developers try it. The trust mechanism decides whether they dare to use it long term. DeepSeek has already pushed the door open with price. What it needs to answer next is whether the scoring rules behind the door can withstand being used as infrastructure.