DeepSeek Opens Search and Fetch APIs at $0.75 per 1,000 Calls

DeepSeek just put a price tag on web search for AI agents: $0.75 per 1,000 successful requests. The company opened its in-house Search and Fetch APIs to external developers, ending the era of treating live web access as a bundled afterthought. For teams building agents and RAG applications, this is now a line item on an invoice.

TL;DR: DeepSeek has opened its in-house Search and Fetch APIs to external developers, priced at $0.75 (5 RMB) per 1,000 successful requests, with both endpoints capped at 5 requests per second per account. Tool calls through Claude Code and Codex remain free until October 31, 2026, when billing begins.

What Exactly Did DeepSeek Launch for AI Agents?

DeepSeek opened its in-house search engine to developers as two plain API endpoints: Search and Fetch. The web search behind DeepSeek’s own chat app is now sold as a standalone service for agents and RAG applications, according to AI/TLDR coverage of the release. That means developers no longer need to scrape results or stitch together third-party providers.

The two endpoints serve different jobs. Search queries the web and returns results; Fetch retrieves a specific page. Both return structured output designed for machine consumption, not human reading.

This launch arrived alongside a model release. DeepSeek posted V4.1-Flash-1017 to Hugging Face on October 17, 2026 under an MIT license, paired with what AI Weekly describes as an open rate card for its Search and Fetch APIs. The same model scores 77.0 on DeepSWE v1.1, per MarkTechPost.

Why does a search API matter this much? Because agents are only as good as the fresh information they can reach.

How Much Does Search and Fetch Cost, and When Does Billing Start?

Each endpoint costs 5 yuan — roughly $0.75 USD, or about 3 PLN — per 1,000 successful requests. The keyword there is successful. Failed calls do not burn your budget, a detail Pandaily highlights in its coverage of the launch.

The timeline matters as much as the number. Tool calls made through Claude Code and Codex remain free until October 31, 2026. PPC Land reports that the $0.75 pricing takes effect from that date, when billing formally begins.

That gives developers a runway to integrate and test before any charge lands. The structure is deliberately simple: one price, two endpoints, no tier maze. Compare that with usage-based search pricing that varies by region, query type, or result count, and the flat rate reads as a deliberate simplification.

For teams processing large volumes, the math is straightforward. Ten thousand successful fetches cost $7.50. A million runs to $750 — real money, but predictable money.

What Do the Search and Fetch Endpoints Actually Return?

The endpoints return structured web results or a page converted to Markdown, per Pandaily. This choice solves a practical problem for anyone who has built RAG pipelines.

Raw HTML is messy. Parsers break, layouts change, and cleaning pipelines rot. Markdown, by contrast, preserves headings, lists, and code blocks in a format that language models ingest cleanly. A fetched page arrives pre-normalized, ready for chunking and embedding.

The split between Search and Fetch maps to how agents actually operate. Search answers “where is the information?” Fetch answers “give me the full document.” An agent might Search for recent announcements, then Fetch the two most promising pages for detailed extraction. Each step bills separately at the same rate.

Kingy AI’s coverage frames the pair as a fresh connection between AI apps and live web content, with practical uses centered on agents and retrieval workflows. In short: the output format is built for pipelines, not browsers.

What Are the Rate Limits and Usage Restrictions?

Both endpoints are capped at 5 requests per second per account. That limit comes straight from Pandaily’s report and is repeated across coverage of the launch.

What does 5 QPS mean in practice? It is generous for a single agent answering questions interactively. It is tight for a batch job crawling thousands of pages in parallel. Teams running high-throughput indexing will need to throttle their workers or distribute load carefully.

A simple calculation shows the ceiling. At 5 requests per second, one account can sustain at most 18,000 requests per hour, or roughly 432,000 per day if running flat-out around the clock. Billing only counts successful requests, so retries on failures carry no direct cost — but rate limits apply regardless.

The per-account framing also matters architecturally. Since the cap is not per API key or per endpoint, splitting workloads across multiple keys under one account will not multiply throughput. Plan your concurrency budget accordingly. For most agent use cases — search, read, synthesize, respond — the ceiling sits well above what a typical conversation needs.

Why Is the Free Access via Claude Code and Codex Significant?

Here is the twist in DeepSeek’s rollout: tool calls through Claude Code and Codex stay free until October 31, 2026, as both PPC Land and Startup Fortune report. Once that date passes, billing begins for those calls too.

This grace period serves two purposes. First, it lets developers wire the endpoints into coding agents and automation tools with zero cost pressure. Testing search quality, latency behavior, and result relevance takes time — and time is free here until the end of October 2026.

Second, it signals where DeepSeek sees demand coming from. Claude Code and Codex are agentic coding environments, tools where an assistant needs to look up current documentation, library versions, and error messages on the fly. Free access there seeds exactly the workflows most likely to convert into paid volume later.

The question PPC Land poses is the right one: what changes once billing begins? The answer, mechanically, is simple — successful calls start costing $0.75 per thousand. Strategically, teams that build dependencies on the free tier now face a buying decision. Budget for it before October 31, 2026, or architect a fallback. Either way, the clock is running.

How Does This Connect to DeepSeek’s V4.1-Flash-1017 Release?

The Search and Fetch APIs arrived alongside a new open model. DeepSeek posted V4.1-Flash-1017 to Hugging Face on October 17, 2026 under an MIT license, according to AI Weekly, pairing the weights drop with an open rate card for its Search and Fetch endpoints. The two releases form a package aimed at agent builders.

The model itself carries a benchmark. MarkTechPost reports that DeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, a software engineering evaluation. An MIT license means developers can download, modify, and deploy the weights without license fees.

Why bundle a model with an API rate card? Because agents need both pieces: a reasoning engine and live web access. By releasing the weights openly and pricing Search and Fetch at $0.75 per 1,000 calls, DeepSeek covers both halves of the stack in one announcement. Developers can run the model themselves and buy the web layer, or combine the two through the company’s API.

The pairing also signals intent. AI/TLDR notes that DeepSeek opened the same in-house search engine behind its own chat app as a plain API for agents and RAG applications. The search capability is no longer locked inside a consumer product.

What Do the Data Terms Mean for Developers?

The endpoints return structured data, which shapes how developers integrate them. According to Pandaily, the new endpoints return structured web results or a page as Markdown. That means Search hands back parseable result sets, while Fetch converts a live page into a format that language models can read directly.

Markdown output matters for RAG pipelines. Instead of writing custom HTML parsers or dealing with broken page layouts, developers receive clean text ready for chunking and embedding. Less preprocessing code sits between the API call and the vector store.

Billing is also tied to outcomes. Both Pandaily and KuCoin describe the price as 5 yuan, roughly $0.75, per 1,000 successful requests. Failed calls do not consume budget in the same way, which reduces the risk of paying for timeouts or blocked fetches.

Developers should still plan for operational work. Activepieces’ coverage focuses on what operational responsibilities come with using the service, a reminder that adding a live web layer to an agent brings reliability questions: retry logic, error handling, and monitoring all remain the builder’s job. The API supplies the data; the plumbing around it stays yours.

How Does the Pricing Compare With the Broader Agent Tooling Market?

The headline number is hard to ignore: $0.75 per 1,000 calls, as PPC Land reports, priced from October 31, 2026. At that rate, an agent performing ten web searches per task could handle thousands of tasks for a single dollar of search spend.

Context matters here. Web search has traditionally been bundled into premium chat subscriptions or enterprise AI platforms rather than sold as a raw, metered endpoint. DeepSeek’s move, as AI/TLDR describes it, exposes the search engine behind its own chat app as a plain API for agents and RAG apps — a consumption-based model rather than a seat-based one.

The pricing is also explicit and public. AI Weekly describes it as an open rate card paired with the V4.1-Flash-1017 release. Developers can calculate costs before writing a line of code: 5 yuan per 1,000 successful requests, with a documented cap of 5 queries per second per account.

For comparison, the free tier is equally concrete. Startup Fortune reports that Search and Fetch are free via Claude Code and Codex until October 31. That gives teams a zero-cost evaluation window before the rate card takes effect.

Who Should Use DeepSeek’s Search and Fetch APIs?

Agent builders top the list. Kingy AI’s coverage frames the APIs as a connection between AI apps and live web content, which is exactly what autonomous agents need when their training data goes stale. Any workflow that asks “what happened this week?” requires a live lookup.

RAG application developers are the second clear audience. AI/TLDR explicitly names agents and RAG apps as the target use cases. Because Fetch returns pages as Markdown, grounding a chatbot in current sources becomes a two-call pattern: search for candidates, fetch and embed the relevant pages.

Automation builders form a third group. Activepieces addresses the automation angle directly, describing how DeepSeek search and fetch is available for automation workflows. Structured output makes the endpoints easier to chain inside no-code or low-code platforms.

Who should be more careful? High-throughput systems. The 5 requests per second cap per account, reported by Pandaily and PPC Land, means large fan-out crawls or parallel agent fleets will need queuing, batching, or multiple accounts managed within the terms. For typical conversational agents, however, 5 QPS is rarely a constraint.

What Should Developers Do Before October 31?

October 31, 2026 is the pivot date. PPC Land reports that DeepSeek priced agent web search at $0.75 per 1,000 calls from October 31, 2026, and that Claude Code and Codex tool calls stay free until then. The free window is a finite resource; here is how to use it well.

A practical checklist:

Review the official API documentation at api-docs.deepseek.com/guides/search, the URL shared in the Hacker News discussion captured by Alto
Prototype a search-then-fetch chain against a real use case while calls are free
Measure your average successful-request volume per task or per user session
Test behavior at the 5 QPS cap to see whether your architecture needs throttling
Validate the Markdown output quality against your parsing and embedding pipeline
Log both successful and failed requests to build an accurate cost forecast
Compare the $0.75 rate card against your current web-access solution’s effective cost
Decide whether V4.1-Flash-1017’s MIT-licensed weights fit your self-hosting plans

Cost modeling deserves special attention. Because billing counts successful requests only, a prototype that retries aggressively could look cheaper in testing than in production — or the reverse. Run realistic workloads before the deadline, not toy queries.

In my opinion, the free Claude Code and Codex period is the real story for small teams: it removes the usual excuse for delaying evaluation. Build the integration now, and the billing change becomes a line item rather than a blocker.

Frequently Asked Questions

What does the $0.75 price actually cover?

Each call to the Search or Fetch endpoints costs 5 yuan (roughly $0.75) per 1,000 successful requests, according to Pandaily. Search returns structured web results, while Fetch returns a page as Markdown — and billing applies to successful requests, not every attempt.

Search and Fetch are now standalone APIs at 5 yuan (~$0.75) per 1,000 successful requests, returning structured results or Markdown pages (Pandaily, KuCoin)
Both endpoints cap at 5 requests per second per account, so high-throughput designs need throttling (Pandaily, PPC Land)
Claude Code and Codex calls stay free until October 31, 2026 — a concrete evaluation window (Startup Fortune, PPC Land)
V4.1-Flash-1017 pairs with the APIs: MIT-licensed weights on Hugging Face, released October 17, 2026, scoring 77.0 on DeepSWE v1.1 (AI Weekly, MarkTechPost)
Targets are agents, RAG apps, and automations, with operational responsibilities like retries and monitoring remaining with the developer (AI/TLDR, Activepieces)

If you build agents, test the endpoints before October 31 and forecast your costs against the $0.75 rate card. The official DeepSeek Search & Fetch documentation is the starting point.