DeepSeek Launches Self-Developed Search Engine API
DeepSeek officially opens the Search and Fetch APIs based on its self-developed search engine, providing real-time web content for Agents and RAG applications.
The Search interface returns a structured list of results based on queries, while the Fetch interface retrieves a single webpage and returns content in Markdown format. Both use the same authentication as the model API, with a base address of api.deepseek.com.
According to the official documentation, the search engine processes hundreds of millions of requests daily, indexes hundreds of billions of webpages, supports multilingual retrieval, and filters low-quality content. Search and Fetch are billed based on successful requests, with a price of $0.75 per thousand successful calls; failed requests, empty results, and timeouts are not charged.
The account-level rate limit is 5 queries per second, with Search and Fetch counted independently. Calls through DeepSeek Harness, Claude Code, or Codex's server-side tools are currently free, with billing set to begin at 00:00 Beijing time on November 1, 2026.
On the same day, the DeepSeek-V4.1-Flash-1017 model update was released, with Terminal-Bench 4.0 improving from 31.2% to 42.3%, DeepSWE from 74.2% to 77.0%, and Agents’ Last Exam from 31.8% to 35.4%. The model is now live on websites, applications, and APIs, with weights open-sourced, and API pricing remaining unchanged.
Developers incur per-call fees when using the search interface, with funds flowing from user accounts to DeepSeek, while model inference costs are calculated separately. Cache hits can reduce model token costs, but search calls are billed independently based on successful attempts.
Source: Public Information
ABAB AI Insight
DeepSeek previously entered the model market with open-source weights and low-cost APIs, with the V4 series offering different cost-performance ratios through Flash and Pro tiers. The company’s self-built search engine supports internal networking capabilities, and this time, Search and Fetch are opened as independent tools for third-party Agents and RAG scenarios.
The search API is priced based on successful requests, separate from model token billing. Calls through tools like Claude Code are free until November 1, after which they will switch to per-call charges, transforming internal service capabilities into external revenue sources. V4.1-Flash-1017 improves several agent and coding benchmarks at the same price while reducing output length and task duration.
Similar paths can be seen among other model providers productizing internal retrieval or tool capabilities, such as some companies selling web scraping or search as additional APIs. DeepSeek is currently in a phase of expanding from pure model supply to a model plus tool stack, with search capabilities becoming a billable component in Agent workflows.
This represents a restructuring of the industry chain: transforming the self-developed search engine from internal support to an external API, while maintaining price competitiveness through model updates. Per-call billing creates stable revenue from high-frequency Agent calls, complementing token billing and changing the revenue structure that solely relies on the scale of model inference.
ABAB News · Cognitive Laws
Opening internal search to the outside equals adding a new billing layer
Success-based billing per call equals making failures free
Higher scores at the same price equal a leftward shift in the cost curve