DeepSeek's Search API and Fetch API are charged per successful request: 5 yuan per 1,000 successful requests. Failures, empty results, and timeouts are not charged. Direct calls to these two APIs have been charged since they launched on October 17, 2026. Server-side calls triggered inside Claude Code, Codex, and DeepSeek Harness also began being charged per request from 00:00 Beijing time on November 1, 2026. If you still budget for November onward as if they were free, you will end up with an extra charge.
Two APIs, the Same Division of Labour and the Same Price
Search searches the web by keyword and returns titles, links, summaries, and publication times, up to 50 results at a time; Fetch fetches a single web page and converts it to Markdown. Both APIs use the same DeepSeek API key, at the official Search endpoint and Fetch endpoint respectively. They are charged per request, not per token and not per returned result. Billing only counts success: a Search call returning 200 with a non-empty result array counts once, and a Fetch call returning 200 with body text retrieved counts once. Empty results, 400, 401, 402, 403, and 404 do not count.
Errors and Rate Limits: Handle Them by Status Code First
Insufficient balance returns 402 — top up before retrying. Deduction takes gifted balance first, then topped-up balance. For 408, 429, and 5xx, wait and retry; a retry that succeeds is billed as normal. A Fetch call returning 404 mostly means the target page cannot be opened, is blocked by the site's policy, or no body text can be extracted; changing the request will not help, and blind retries only waste time. Both APIs are limited to 5 QPS per account, and Fetch has an additional per-domain rate limit. Batch fetching should be queued first, with concurrency controlled.
Work Out the Cost Before Running Jobs in Bulk
One search plus fetching 5 web pages is 6 successful requests in total, costing 0.03 yuan in tool fees. The real bulk of the cost comes afterwards: the Markdown fetched back has to be fed to a model, and token fees are charged separately. The APIs have no pagination field, so sending the same keyword again will only return the same earlier results repeatedly, and each send is billed. To turn off this cost in Claude Code, add WebSearch to the denied permissions in settings. First test with a small batch, confirm the results are usable, and only then scale up. Cheap tool fees do not mean retries are free.
