DeepSeek Search & Fetch API: requests, limits and billing
Choose direct Search/Fetch requests or DeepSeek server-side tools. Plan shared QPS, successful-request costs and the November 1 billing switch.
Direct requests: search first, fetch selectively
Keep two account queues and a domain-aware Fetch queue
The November switch applies to server-side tool calls
Choose who controls retrieval
DeepSeek opened its Search and Fetch APIs on October 17, 2026. For an agent workflow, the useful distinction is who decides when retrieval runs: your application can call the endpoints directly, or a configured client can use DeepSeek server-side tools. Search returns structured web results; Fetch turns one page into Markdown. This is a documentation review dated October 17, not a service test. We made no API requests, installed no client and measured no search quality, latency or usage.
| Your requirement | Route to inspect | Operational decision |
|---|---|---|
| Your code chooses queries, URLs and when to stop | Direct JSON requests to Search and Fetch | Own the queue, result checks, citation record and request budget. |
| An existing Harness, Claude Code or Codex workflow uses the DeepSeek API | Its configured DeepSeek server-side tools | Confirm actual provider routing and tool usage before applying the temporary billing exception. |
| You only need a known, authorized document | Evaluate whether web retrieval is needed at all | Avoid adding repeated searches to a deterministic document workflow. |
Direct requests: search first, fetch selectively
The official guide documents POST /v1/tools/search and POST /v1/tools/fetch at https://api.deepseek.com, using the existing DeepSeek API bearer authentication and JSON bodies. Search accepts a query; Fetch accepts a URL. Use the official reference for complete schemas. Selecting a result for further reading should depend on relevance and source quality, not simply its position. Fetch remains subject to the target site’s access policies; it is not a way to bypass restricted pages.
| Step | Illustrative JSON body | Documented response |
|---|---|---|
| Find candidate sources | {"query":"release notes site:example.com","count":5,"need_content":false} | A results array with URLs and structured result fields; an empty array can be a normal successful search. |
| Read a selected page | {"url":"https://example.com/article"} | title and content; content is Markdown. This placeholder is not a tested retrievable page. |
The November switch applies to server-side tool calls
The announcement sets the switch at October 31, 2026, 16:00 UTC; the pricing page expresses the same instant as November 1, 2026, 00:00 Beijing time. The temporary free treatment is specifically for Search/Fetch server-side tool calls made when using the DeepSeek API through DeepSeek Harness, Claude Code or Codex. Direct API requests have published per-request prices. Do not extend the temporary exception to all direct usage, or infer eligibility from a Claude or OpenAI subscription. Confirm which provider and billing account the workflow actually uses before budgeting.
Budget successful non-empty requests, then add model tokens
The guide prices Search and Fetch separately at $0.75 per 1,000 billable requests each. Billable means HTTP 200 plus a non-empty Search results array or actual Fetch page content. Failed requests, empty searches and timeouts are not billed. Search count changes the number of results, not the price of that request. Model-token charges are separate. The calculation below is an illustrative budget, not an observed bill or a claim that every request succeeds.
| Budget item | Assumption | Calculated amount |
|---|---|---|
| Search | 10,000 successful non-empty requests × $0.75 / 1,000 | $7.50 |
| Fetch | 30,000 successful requests with content × $0.75 / 1,000 | $22.50 |
| Retrieval subtotal | 40,000 billable endpoint requests | $30.00 |
| Model and other costs | Input/output tokens, hosting and any client charges | Additional; not estimated here |
Measure amplification before increasing volume
For a proposed research workflow, write down searches per task, selected pages per search and repeated visits to the same page. Keep a separate counter for each billable endpoint outcome. Where rights, freshness and application requirements permit, reusing a previously retrieved page can avoid redundant work. A ten-result search is still one request; fetching all ten pages introduces ten further requests. Set an application budget and stop condition, and retain enough redacted usage detail to reconcile it with provider billing. No such workload was executed for this guide.
Handle failure without an unlimited retry loop
DeepSeek instructs callers to wait and retry HTTP 408, 429 and 5xx errors. Our suggested policy adds bounded attempts and backoff so retries cannot create an uncontrolled queue. A 429 may come from an account limit or Fetch’s domain limit. Other 4xx responses require investigation or a changed request rather than blind retry: 400 concerns the request, 401 access, 402 balance, and 403/404 can concern security policy or unavailable usable content. Stop and surface account or billing issues to the responsible operator; changing keys is not a rate-limit solution.
Keep source evidence separate from model instructions
Treat retrieved text as source material. Preserve the URL and retrieval date, inspect whether the page supports the intended statement, and keep conflicting evidence visible. Instructions embedded in a retrieved page should not acquire authority to change the workflow, send a message or access private data. Search availability does not establish permission to republish a page, and an attractive summary does not establish factual accuracy. Use the linked small-task evaluation guide to define acceptance criteria before authorizing a live trial.
Flash-1017 is migration context, not a prerequisite benchmark
The same announcement introduces DeepSeek-V4.1-Flash-1017 and recommends DeepSeek Harness v0.2.1-rc.1 or later for that model. The pricing page maps deepseek-flash to Flash-1017. Review the client and model routing if you maintain a DeepSeek coding workflow, but keep that migration separate from a decision to call standalone retrieval endpoints. We did not install Harness or validate vendor performance claims.
A review checklist for the workflow owner
Before enabling retrieval, record the chosen route, provider account, endpoint queues, domain throttling, retry cap, source-retention rules and maximum requests per task. Mark the billing-switch instant in UTC, calculate a post-switch budget and identify who can stop the workflow. Then define one authorized evaluation with known sources, missing-page and empty-result cases, and a way to inspect citations. For help comparing the surrounding agent framework or review requirements, use the workflow-building guide or contact uAgentKit with those constraints.
Continue your comparison
Plan agent workflow building →
Define a small agent evaluation →
Sources supporting this guide
- DeepSeek: October 17 release and Search/Fetch opening
DeepSeek · api-docs.deepseek.com · Read · 2026-10-17
- DeepSeek API change log: October 17, 2026
DeepSeek · api-docs.deepseek.com · Read · 2026-10-17
- DeepSeek Search & Fetch: requests, billing, errors and rate limits
DeepSeek · api-docs.deepseek.com · Read · 2026-10-17
- DeepSeek Models & Pricing: per-request charges and billing transition
DeepSeek · api-docs.deepseek.com · Read · 2026-10-17
Research method & original evidence · Who operates uAgentKit · Commercial disclosure