DeepSeek Search and Fetch End Free Tier, Billed at $0.75 per Thousand Calls
The two web tool APIs, DeepSeek Search and Fetch, end their free tier starting today and switch to billing by the number of successful requests, at the same unit price of $0.75 per thousand calls. Billing only counts requests that return HTTP 200 and where Search returns non-empty results or Fetch actually retrieves page content; failures, empty searches, and timeouts are not billed, and the cost is independent of the number of returned items, tokens, and whether need_content is enabled. Rate limits are calculated per account: Search is 100 QPS and Fetch is 25 QPS, with Fetch also having a per-domain QPS cap; server-side tool calls initiated via DeepSeek Harness, Claude Code, and Codex are subject to the same rules.
Why read this
AI score 62
The original article lists the billing criteria, success determination, and rate-limiting rules one by one, so readers can estimate the actual cost of an agent looping through web tool calls.
Topics
Body · AI translation
DeepSeek Search / Fetch, two independent web tool APIs invoked as server-side tool calls through DeepSeek Harness, Claude Code, Codex and similar products, are ending their free tier as of today and will begin charging by the number of successful requests, at the same unit price of $0.75 per thousand calls.
# What each one does
Search POST /v1/tools/search: give it a query and it returns a list of results; by default you can request 10, up to a maximum of 50. Each result has a title, URL, a short snippet (for humans to read), a long summary (for the model to read), optional full-text content, site name/icon, date, an authority score of 0–4, and a rerank_score of 0–3 (the document says ≥1.5 counts as high quality). It supports operators such as site:, -site:, exact phrases, lang:, after: / before:; you can also pass the user's original question, IP-based localization, and need_content.
Fetch POST /v1/tools/fetch: give it an HTTP(S) URL and it returns the title plus the body in Markdown. It is subject to the target site's availability and access policy; a rejection by the security policy is 403, and failure to extract usable content is 404.
https://t.co/KTvGSEbiTs
# How billing works and what counts as "success"
Fee = number of successes × unit price, independent of the number of results returned, tokens, or whether need_content is enabled; count only adjusts the number of results and does not add to the price. The billable conditions are stated clearly: it must be HTTP 200, and Search must return non-empty results while Fetch must actually retrieve page content. Failures, empty searches, and timeouts are not billed. Charges are deducted first from gifted credit, then from the topped-up balance; if the balance is insufficient, you get a 402.
How to convert: a thousand Searches or a thousand Fetches are $0.75 each; if an Agent loop makes 10,000 successful Searches in a day, that is about $7.5, not including model tokens.
# Rate limits (calculated separately from model concurrency)
Counted per account, not per API Key: Search 100 QPS, Fetch 25 QPS; they do not eat into each other's quota. Fetch also has a per-domain QPS cap, and exceeding it likewise returns 429. For 408 / 429 / 5xx you can wait and retry; for other 4xx, change the request before sending it again, and don't retry blindly.
# Relationship with Harness / Claude Code / Codex
The documentation states clearly: when going through these products, server-side tool calls will hit Search and Fetch anyway. The free period ends at midnight Beijing time on 2026-11-01; today is already 11-02, which amounts to "an extra tool bill on top of the model bill." Calling these two APIs directly yourself follows the same rules.