DeepSeek launched its own search engine and API on 17 October 2026, moving from third-party providers to a proprietary crawling and indexing stack. The new infrastructure supports hundreds of billions of web pages and handles hundreds of millions of search requests per day. The launch impacts LLM developers seeking low-cost retrieval and webmasters managing AI crawler traffic.
Is this for you?
If you develop LLM applications, DeepSeek's new in-house API offers lower-cost, Markdown-ready data retrieval starting at $0.75 per 1,000 successful requests.
- If you are an LLM developer, the Fetch endpoint provides a more direct path to LLM-ready data in Markdown format.
- If you are a webmaster, you should update firewall rules to recognize the DeepSeekBot/1.0 user agent.
- If you are an AI SEO practitioner, note that DeepSeek filters GEO as a type of spam.
- Skip this if you are looking for specific traffic or click impact metrics as these were not provided.
What you can do now
- If you use the DeepSeek API, check if your account can exceed the 5 queries per second limit.
- If you use Claude Code or Codex, use the free access to these endpoints before 00:00 Beijing time on 1 November 2026.
- If you manage a website, update your firewall to recognize the DeepSeekBot/1.0 user agent string.
Key figures
- Launch Date
- 17 October 2026
- API Cost
- $0.75 per 1,000 successful requests
- Read Throughput
- 8 TB/s
- Storage Capacity
- 20 PiB
- Query Rate Limit
- 5 queries per second per account
- User Agent
- DeepSeekBot/1.0
If you currently use third-party search APIs or standard search tools without native Markdown fetching, DeepSeek’s new Fetch endpoint provides a more direct path to LLM-ready data. (Markdown: a lightweight markup language used to format text and structure data for easy reading and ingestion.)
DeepSeek Search and Fetch API Launch
DeepSeek released two specific endpoints on api.deepseek.com: a Search endpoint (POST /v1/tools/search) for web results and a Fetch endpoint (POST /v1/tools/fetch) for downloading pages. DeepSeek says the Fetch endpoint returns content in Markdown format, which simplifies the data ingestion pipeline for developers by handling the conversion automatically.
DeepSeek transitioned from using the Bocha search API in 2025 to this in-house stack by September 2026. By owning the infrastructure, DeepSeek can offer a low-cost API for search-augmented LLM applications while maintaining direct control over the ranking and spam detection logic.
API Pricing and Query Limits
DeepSeek sets the cost for both the Search and Fetch endpoints at $0.75 per 1,000 successful requests. DeepSeek does not bill for empty results, failures, or timeouts, which helps maintain predictable costs for high-volume applications.
Query rate limits are set at 5 queries per second per account. For specific use cases, DeepSeek provides free access to these endpoints for calls made from DeepSeek Harness, Claude Code, and Codex until 00:00 Beijing time on 1 November 2026.
Hardware Throughput and Storage
DeepSeek’s infrastructure is built on a cluster of 180 storage nodes using a proprietary 3FS file system. (3FS: a proprietary file system used by DeepSeek for their storage cluster.) The hardware setup includes 2,880 NVMe SSDs providing a total capacity of 20 PiB. (NVMe SSDs: high performance solid state drives that use the non volatile memory express connection standard.)
Measured read throughput for this cluster reaches 8 TB/s. (throughput: the rate at which data is processed or moved through a system over a period of time.) This massive hardware throughput is required to support the scale of a multi-billion page index, allowing the engine to handle the hundreds of millions of daily requests DeepSeek reports.
Treatment of GEO as Spam
DeepSeek categorizes Generative Engine Optimization (GEO)—the practice of editing content so AI engines cite it—as a type of spam. The company explicitly states that its index filters out content identified as GEO spam.
Search operators familiar from Google are supported within the API, including site:, -site:, -term, exact phrases in quotes, lang:, after:, and before:. These operators allow for more granular control when querying the DeepSeek index.
LLM Developers and RAG Pipelines
LLM application developers building search-augmented agents should switch to the Fetch endpoint to retrieve documentation directly in Markdown. This change reduces the need for custom parsing logic in the RAG pipeline. (RAG pipeline: a system architecture that retrieves relevant information to provide context for large language model responses.)
Search Infrastructure Engineers
Search infrastructure engineers evaluating large-scale indexing can use the 3FS and NVMe throughput figures as a benchmark for high-performance search clusters. The 8 TB/s throughput provides a concrete performance target for high-density storage architectures.
AI SEO Practitioners
AI SEO practitioners evaluating content strategy should note that DeepSeek filters GEO as spam to avoid content manipulation. This means that traditional GEO tactics may lead to index filtering rather than increased visibility.
Webmasters and Bot Identification
Webmasters managing crawler traffic should update firewall rules to recognize the DeepSeekBot/1.0 user agent. (user agent: a string identifying the software and operating system making a request to a web server.) DeepSeek’s training-data summary for the EU confirms that the crawler respects robots.txt rules.
Example: Diego’s Research Agent
Diego runs a research agent that summarizes technical documentation from the web. He integrates the DeepSeek Fetch endpoint to pull pages into his system, noting that the Markdown output saves him the time of writing a custom parser.
Diego monitors his costs based on the $0.75 per 1,000 requests rate and helps ensure his agent does not exceed the 5 queries per second limit. If Diego’s requirements increased to 50 queries per second, he would be unable to use the standard endpoint and would need to negotiate a higher limit or scale his architecture.
Practical Applications
| Application | User |
|---|---|
| Cost-modeling a RAG pipeline | Growth Engineers |
| Identifying and blocking the crawler | Webmasters | ight
| Implementing Google-style queries in an AI app | LLM Developers |
| Benchmarking Markdown extraction | Data Engineers |
| Monitoring GEO impact | AI SEO Specialists |
What the Fetch Endpoint Replaces
The DeepSeek Fetch endpoint replaces the Bocha API for search results, providing DeepSeek with direct control over the crawling and indexing stack. By moving in-house, DeepSeek eliminates the dependency on a third-party provider for its core search functionality.
Unconfirmed Index Metrics
DeepSeek did not provide a specific measurement of its total index size, the effect on traffic or clicks, or the specific regions and languages covered by the new engine. The Chinese and English privacy policies do not specify whether any search traffic still routes to external providers as of October 2026.
DeepSeek: what to check, compare and watch
- Perplexity API
- Developers can compare the cost and markdown output of DeepSeek's Fetch endpoint against the Perplexity API. LLM application developers: compare the $0.75 price point against the current Perplexity API pricing to determine the best cost-to-performance ratio.
- Exa API
- DeepSeek is listed as a search provider alongside Exa in the deepseek-harness tool. LLM application developers: switch to the Exa API if the DeepSeek search results do not meet the required accuracy for the specific use case.
- NVMe SSD
- DeepSeek's infrastructure uses 2,880 NVMe SSDs to support its internal search index. Count: 2,880. Search infrastructure engineers: monitor the performance of NVMe SSDs in high-throughput search clusters to benchmark against DeepSeek's architecture.
- Common Crawl
- DeepSeek uses Common Crawl as the starting list of pages for its own crawler. Search infrastructure engineers: verify the completeness of Common Crawl data before initiating a new crawl of the web index.
- Server logs
- The DeepSeekBot identifies itself in server logs, allowing infrastructure engineers to monitor its activity. User agent string: Mozilla/5.0 (compatible; DeepSeekBot/1.0; +https://www.deepseek.com/). Search infrastructure engineers: update firewall rules to whitelist or throttle the DeepSeekBot user agent based on crawl frequency.
- SEO spam
- DeepSeek categorizes GEO as a type of content manipulation similar to SEO spam. AI SEO practitioners: compare current content tactics against DeepSeek's definition of SEO spam to avoid index filtering.
DeepSeek: questions the original report leaves open
- What changed, and where — which regions, languages or devices?
- The original report did not say what regions, languages or devices changed.
- What effect does it have on traffic or clicks, and what is the evidence?
- The original report did not say what effect it has on traffic or clicks.
- What do site owners need to change — markup, settings or content?
- The original report did not say what site owners need to change.
- Can it be measured or reported, and where?
- The original report did not say where it can be measured or reported.
- How is the V4 authority token calculated for different domains?
- The original report does not say.
Key terms
- GEO
- Generative Engine Optimization—editing web content so that AI answer engines cite it.
- 3FS
- DeepSeek’s proprietary file system used in their storage cluster.
- Fetch Endpoint
- An API endpoint that downloads a web page and returns it in Markdown format.
- V4 Authority Token
- A technical token that matches the authority field returned by the Search API.
Original reporting: DeepSeek Search Engine (DEJAN), on 18 October 2026.