DeepSeek API Web Search No Longer Costs Extra

Share

From November 18 at 12:00 Beijing time, the DeepSeek API no longer charges a tool fee when the model runs web search or page fetching on its own, so you pay only for tokens. Direct calls to the Search and Fetch APIs are still billed per request. The article lists what is free and what still costs money, how Claude Code connects through DeepSeek, and the deepseek-flash peak and off-peak prices.

Machine translation. Read the original

DeepSeek has changed its billing. Starting November 18 at 12:00 Beijing time, web search and web page fetching that the model starts on its own during a conversation are no longer charged separately. You only pay for the tokens the model uses. When DeepSeek opened these web tools on November 13, it also charged a separate tool-call fee on top of token costs. That fee is now gone. If you connect an AI agent like Claude Code to DeepSeek, letting it search the web costs you less from today.

What's free now and what still costs money

How it's used How it's billed now
Declaring deepseek_web_search in a conversation (web search and deep-thinking search) Search is free. You pay only for tokens
Declaring deepseek_read_url in a conversation (reading a link the user sends) Fetching is free. You pay only for tokens
Calling the web_search tool through the Anthropic-format interface Search is free. You pay only for tokens
Calling the web_search tool through the Responses API Search is free. You pay only for tokens
Your own program calling the Search API directly (/v1/tools/search) Still charged, CNY 5 per 1,000 successful requests
Your own program calling the Fetch API directly (/v1/tools/fetch) Still charged, CNY 5 per 1,000 successful requests
File parsing and OCR Still charged, CNY 1.5 per 1,000 pages

Simple rule: if the model decides for itself whether to search, there is no extra charge. If your program skips the model and uses the API directly as a search engine, you pay per call.

What this means for people using AI agents

DeepSeek has published a setup guide for Claude Code. Set ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic and the model to deepseek-flash. Claude Code's built-in Web Search will then use DeepSeek's search. DeepSeek's docs say the search request itself is free. But the search results still go back to the model to be summarized. That step is one more model request, so it adds token costs.

Web tools currently require the deepseek-flash model. Its pricing has peak and off-peak rates. Off-peak is half the peak price:

deepseek-flash (per million tokens) Off-peak Peak
Input, cache hit CNY 0.02 CNY 0.04
Input, cache miss CNY 1 CNY 2
Output CNY 4 CNY 8

Peak hours are weekdays, Beijing time, from 9:00 to 12:00 and from 14:00 to 18:00. Weekends and public holidays are off-peak all day. For jobs that search a lot of material, run them in the evening or on the weekend.

If you use DeepSeek Harness, see The Community Has Fixes for DeepSeek Harness's Biggest Headaches for common questions and community plugins.

Ask your AI agent to check

Starting November 18, 2026 at 12:00 Beijing time, the DeepSeek API no longer charges separately for web search and web page fetching that the model starts on its own during a conversation. You only pay for tokens. Direct calls to the Search and Fetch APIs are still charged per request. Please help me check whether I can save this money:

1. Find every place in this project and my AI agent configuration that uses DeepSeek (environment variables, config files, code). Put them in a table with the model used, the API format used, and whether it calls web search or fetch.
2. Mark which ones are "the model searches on its own" (now free) and which are "the program calls /v1/tools/search or /v1/tools/fetch directly" (still charged).
3. For the second group, if they could be changed so the model searches on its own, tell me how to change them and what would be different. For any that can't be changed, explain why.
4. Read only. Don't change anything, and don't make any real API requests. I will confirm before we make changes.

Notes

  • Search is not completely free: Only the tool fee is waived. The tokens the model uses to read search results and write the summary are still charged. The more you search, the more tokens you use.
  • Only deepseek-flash works: Web tools require this model. The old name deepseek-v4-flash still works, but requests are routed to the new Flash model and billed at Flash prices.
  • Prices may change: DeepSeek reserves the right to change prices. Check the pricing page before running large batch jobs.
  • Turn on safety settings if you run a public product: DeepSeek requires developers who offer services to the public to turn on the model's safety parameters. You are also responsible for content review and AI labeling.

References

Done with this one? Mark it to track your path.Sign in
Save it for laterSign in
Questions about this piece?Ask AI about this

Comments (0)

No comments yet. Start the conversation.

Sign in to join the discussion.Sign in