DeepSeek launches Search and Fetch APIs for AI agents at $0.75 per 1,000 calls

What DeepSeek launched

DeepSeek has opened two standalone web APIs aimed at AI agents and retrieval-augmented generation apps. The Search and Fetch guide in its API documentation, live on 17 October 2026 with a banner announcing the launch, describes a Search API that finds web pages for a query and returns structured results, and a Fetch API that retrieves a single page and returns its content as Markdown.

DeepSeek says Search is powered by its own in-house search engine. Both are plain JSON endpoints at api.deepseek.com, under /v1/tools/search and /v1/tools/fetch, and they use the same bearer API key as the company's chat models. That means a developer who already calls DeepSeek's models can add live web access without signing up for a separate search provider.

The move makes DeepSeek, best known for low-cost open-weight models, a supplier of the plumbing that agents need as well as the model. It also fits a pattern we have tracked in its tooling, including DeepSeek Harness, its open agent app, which gained desktop builds in late September.

How the Search API works

A Search request needs only a query of 1 to 128 characters. Optional fields set how many results to return, from 1 to 50 with a default of 10; pass the user's original question to improve query understanding; supply an IP address to localise results; and ask for each page's full content. The query keeps standard search syntax, and the guide lists site and minus-site filters, excluded terms and phrases, exact phrases, a lang filter such as lang:zh,en, and after and before date filters.

Each result carries a title, URL and site name, a short snippet of roughly 100 to 200 characters meant for display, and a longer summary of roughly 600 to 800 characters that DeepSeek says is intended to be read by a model. Results also include a publication date when known, an optional full-text field, a rerank score from 0 to 3 that blends relevance, authority and freshness, and an authority tier from 0 to 4. DeepSeek says scores of 1.5 or more are high-quality results and tiers of 3 or more are highly authoritative sources.

Those two scores are the most agent-friendly part of the design. An agent can drop low-scoring pages before spending tokens on them, or prefer high-authority sources for factual questions, without running its own ranking step.

How the Fetch API works

Fetch takes a single URL of up to 8,192 characters. It must be HTTP or HTTPS, include a host and contain no username or password. The response is just a page title and the page content converted to Markdown, the format most agent frameworks already feed to models.

The guide is frank that fetching depends on the target site. A 403 means DeepSeek's security policy rejected the request, and a 404 covers unavailable URLs, URLs blocked by site policy, errors from the target site, unsupported content types and pages with no extractable Markdown. Errors use the OpenAI-compatible error format. DeepSeek tells developers to retry timeouts, rate-limit errors and server errors after a pause, to top up before retrying a 402 insufficient-balance error, and not to retry other client errors unchanged.

What it costs and how fast you can call it

On DeepSeek's pricing page, Search and Fetch each cost $0.75 per 1,000 successful requests. The Chinese-language guide lists 5 yuan per 1,000. A request is billed only when it returns HTTP 200 with a result: Search must return a non-empty list and Fetch must return page content. Failed requests, empty searches and timeouts are free, and the number of results requested does not change the price.

In practice, an agent that runs ten searches per task would pay $0.75 for a hundred tasks' worth of searching, before the cost of the model tokens it spends reading the results. For comparison, DeepSeek's current Flash model, DeepSeek-V4.1-Flash, is listed at $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, with rates doubling in two weekday peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC.

Throughput is the tighter constraint. Each account gets 5 requests per second for Search and 5 for Fetch, counted separately and across all of the account's API keys, independent of model concurrency limits. Fetch also has a per-domain limit. Requests over either limit receive HTTP 429. Five per second is enough for a single coding agent or a small app, but a busy multi-user product would hit it quickly.

The free period for coding agents

The billing note contains the most interesting detail. DeepSeek's API can be called in OpenAI and Anthropic formats, so tools such as Claude Code and Codex can be pointed at DeepSeek models. The guide says that when people use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit the Search and Fetch APIs, which suggests their built-in web search is now being served by DeepSeek's own engine when it runs against DeepSeek's API.

Those calls are free for now. Per-request billing for them starts at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October. Developers who run long agent sessions against DeepSeek through third-party coding tools should expect a new line on their bill from that date, even if they never call the Search API directly.

Other DeepSeek changes on the same day

DeepSeek also published DeepSeek Harness v0.2.1-rc.1 on GitHub on 17 October, the first release candidate in the 0.2.1 series. Its notes list installing plugins from custom index sources, a button that lets the agent create a plugin, an experimental Git Worktrees plugin, masking of keys and tokens found in sent messages, and an experimental Claude Code Mods compatibility layer that the developers say is meant to test whether that API is broadly a subset of Harness plugins rather than to offer full compatibility.

Separately, Chinese tech site IT Home reported that a bot account on X spotted a model entry named deepseek-v4.1-flash-1017 in DeepSeek's API model list that was quickly removed. The suffix matches DeepSeek's dated snapshot names, but the company had not commented, and its pricing page still maps deepseek-flash to DeepSeek-V4.1-Flash, released on 10 September according to its change log.

What it means for developers

For developers, the appeal is one key, one bill and one vendor for both the model and its web access. Results arrive pre-summarised in a length designed for models, with ranking signals attached, which can cut the glue code that agent builders usually write around a general-purpose search engine. Teams comparing private AI stacks, such as Brave's encrypted Leo chat running GLM 5.3 Flash, now have another option where search and model come from the same company.

It also raises the usual sourcing questions. A model reading search summaries is one step further from the page than a model reading the page itself, so agents that need exact quotes should use Fetch or the full-content option. Open research agents such as OneSearch-VL show how much a deep-research pipeline depends on the quality of the underlying search results.

Limits and open questions

DeepSeek has not published details about its search index: how large it is, how often it is refreshed, which languages and regions it covers best, or how it decides authority tiers and rerank scores. The guide does not say whether queries or fetched pages are stored, used for training, or processed only in China, which matters for companies with data-residency rules.

The documentation also leaves open whether the 5 requests-per-second limits can be raised, and whether the free period for Claude Code and Codex tool calls will be extended. AI Lookout read the documentation and pricing pages but did not run test queries, so we cannot yet say how the results compare in quality with established search APIs.