DeepSeek opens Search and Fetch API to agents: its internal engine at $0.75 per thousand calls
On October 17, DeepSeek quietly turned an accessory into a product
On October 17, 2026, DeepSeek published two announcements on the same day. The first made the headlines: V4.1-Flash-1017, a MoE model of roughly 552 billion parameters in open weights, credited with 77.0 on DeepSWE v1.1. The second went almost unnoticed, yet it's the more interesting one: the company opened its Search and Fetch APIs to everyone — the internal search engine its agents were already using behind the scenes, as reported by Marktechpost.
Concretely: two REST endpoints, billed at $0.75 per 1,000 successful calls (about 5 ¥), powered by DeepSeek's in-house search engine. And until November 1st, these calls are even free when they go through DeepSeek Harness, Claude Code, or Codex.
Why does this matter? Because models are becoming commoditized — V4.1-Flash-1017's weights are MIT-licensed — and value is shifting to the layers around them: orchestration, inference, and now search. By opening up its search layer, DeepSeek isn't just selling a brain to your agents. It's also selling them eyes.
The Essentials
- Search and Fetch are two public APIs:
POST https://api.deepseek.com/v1/tools/searchand/v1/tools/fetch, Bearer authentication, powered by DeepSeek's in-house search engine. - Price: $0.75 per 1,000 successful requests, i.e., $0.00075 per call, for both Search and Fetch (October 2026, verify on api-docs.deepseek.com).
- Free through October 31 inclusive for server-side calls via DeepSeek Harness, Claude Code, and Codex. Billing starts on November 1, 2026 at 00:00 Beijing time.
- Cap of 5 requests/second per account: designed for agents, not for building a consumer search engine.
- V4.1-Flash-1017 shipped the same day: 77.0 on DeepSWE v1.1 (vendor benchmarks, independent ones not yet published), roughly $0.20 per task versus ~$0.50 previously.
- The strategic read: the weights are MIT, the search engine is not. DeepSeek opens up what is becoming commoditized and keeps what differentiates.
Recommended Tools
| Tool | Primary use | Price (October 2026) | Best for |
|---|---|---|---|
| DeepSeek Search API | Web search for agents | $0.75 / 1,000 successful requests | Agents that need up-to-date sources |
| DeepSeek Fetch API | Extracting content from a URL | $0.75 / 1,000 successful requests | Targeted reading of pages already identified |
| DeepSeek Harness | Code agent orchestration | Included (tokens billed separately) | DeepSeek ecosystem |
| Claude Code / Codex | Coding agents with server-side search | Search/Fetch free until Nov 1, 2026 | Teams already equipped with these harnesses |
For the generation side, our guide to free AI APIs lists what's available beyond DeepSeek.
Search and Fetch — what exactly are they?
Two REST endpoints on api.deepseek.com, an in-house search engine, and a clear philosophy: giving agents programmatic access to the web, without a browser or scraping.
| Search | Fetch | |
|---|---|---|
| Role | Discover | Extract |
| Input | Query (query) |
Known URL |
| Output | Ranked and scored results | Page content |
| Typical use case | "What's new about X?" | "Read this specific doc" |
Search is the discovery endpoint. You send a query, a number of results (count), optionally a location (localization), and a need_content flag to request — or not — the full content of the pages. An ori_question parameter also lets you pass along the user's original question.
Fetch does the opposite: you already have the URL, and it fetches and extracts the content.
The Search response is designed for an agent pipeline. Each result includes a rerank_score, a title, the url, a snippet, a summary, the content (if requested), the site_name, a date in ISO format, and an authority field. Two fields deserve a closer look:
rerank_score: a relevance score computed on DeepSeek's side. Your agent can discard weak results without spending a single LLM token.authority: a measure of how reliable the source is, for weighing an official site against an anonymous blog.
It's the kind of detail that gives away the design: this API wasn't built for humans who search, but for models that read.
A call looks like this:
curl -X POST https://api.deepseek.com/v1/tools/search \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "tarifs api deepseek 2026", "count": 5, "need_content": false}'
The price: $0.75 per thousand, free until November 1st — and a cap to watch
$0.75 for 1,000 successful requests, Search as well as Fetch. Failed calls aren't billed: the official documentation does indeed say "successful requests".
Until October 31st inclusive, these calls are free in one specific case: when they go server-side through DeepSeek Harness, Claude Code, or Codex. If you use Claude Code with a DeepSeek backend, every search your agent runs is already hitting these APIs — perhaps without you even knowing it. From November 1st, 2026 at 00:00 Beijing time, billing kicks in, as AI Weekly points out.
Do the math. A chatty agent that fires off 200 searches a day, 30 days a month, burns through 6,000 calls — that's $4.50 per month. Search isn't where your budget goes. The real bill is in tokens: every result that enters the context is billed input at between $0.15 and $0.30 per million on deepseek-flash, depending on the time of day.
And that's the whole logic behind the pricing. Search is given away almost for free because it fills the main meter: the context of a 1-million-token model.
The real issue is the cap of 5 QPS per account. For an interactive agent, that's more than enough. For a product that would have thousands of users searching in parallel, it's a dealbreaker. DeepSeek is targeting agents, not search applications.
The Accessory That Became a Product: The Real Strategic Read
DeepSeek isn't selling yet another model. It's selling a complete stack: the brain (open weights), the hands (harness), and now the eyes (Search/Fetch). That's exactly what Dealroom headlined: DeepSeek "moves up the stack."
Look at the cost structure. The V4.1-Flash-1017 weights are under MIT, downloadable from Hugging Face and ModelScope. Anyone can run the model. But the search engine that feeds it? Proprietary. This is inverted open-core: open up what's becoming commoditized, keep what differentiates — the index, the reranking, the authority scoring.
The economic loop is elegant, slightly perverse. Every search enriches the context, the context consumes tokens, DeepSeek bills for the tokens. An API at $0.00075 per call isn't a revenue source: it's a token-consumption generator. Search is the fuel, not the product.
This move fits into a clear October 2026 trend: the web is adapting to agents. ToolCUA shows Computer Use agents learning to choose between GUI and API — and the API almost always wins: faster, more reliable, cheaper. Salesforce Headless 360 pushes the same logic into the enterprise: a CRM without a browser, all API, designed to be consumed by agents. DeepSeek Search is the open-web version of the same move: rather than letting your agents click around in a browser, give them an endpoint.
On the competition front, OpenSeeker-v2 shows that open-source is already attacking the monopoly of industrial search agents. DeepSeek is taking a different position: neither open-source nor closed — open-API at cost. Two strategies to break the same lock.
One caveat, though. The index is a black box: you control neither the corpus, nor per-language freshness, nor the authority criteria. And the day DeepSeek adjusts the pricing — it was free just a month ago — your agent is already architected around it. Abstract the search layer behind your own interface if you ever plan to switch providers.
V4.1-Flash-1017 in 90 seconds: the fuel behind the free tier
The model isn't the subject of this article — we cover it in detail in our analysis of V4.1-Flash MIT 552B — but it needs to be framed, because the economics of Search/Fetch rest entirely on it.
The vendor figures, published on October 17, 2026: 77.0 on DeepSWE v1.1 versus 74.2 for the previous version, ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) in DeepSeek's table. Terminal-Bench 4.0 goes from 31.2 to 42.3, ProgramBench from 20.3 to 29.1. A word of caution, however: these scores have not yet been independently verified, and Claude Opus 5 retains the advantage on HLE, Terminal-Bench 4.0, ProgramBench and NL2Repo-Bench.
| deepseek-flash (V4.1-Flash-1017) | Value |
|---|---|
| Context | 1M tokens |
| Max output | 384K tokens |
| Vision | ✓ |
| Concurrency | 2,500 |
| Input (cache miss) | $0.15 off-peak / $0.30 peak per 1M |
| Input (cache hit) | $0.003 / $0.006 per 1M |
| Output | $0.60 / $1.20 per 1M |
| Peak hours | 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday |
October 2026 pricing, verify on api-docs.deepseek.com.
Announced result: roughly $0.20 per DeepSWE task, versus ~$0.50 before the update. The claim of a "Pareto improvement" — better performance and cheaper, according to the official release note — owes a lot to shorter outputs. At $0.60–1.20 per million output tokens, every token saved counts fourfold.
A useful detail for European teams: peak hours correspond to Chinese office hours. A good chunk of your working day in Paris or New York falls during off-peak. Schedule your batch processing accordingly.
To situate this model against the competition, our comparison of the best LLMs for coding has been updated accordingly.
Hooking Search/Fetch into your agent: three paths
The simplest option: do nothing, if you're already going through one of the three supported harnesses.
Path 1 — DeepSeek Harness. Version v0.2.1-rc.1 or higher is recommended to take advantage of the new model. Server-side search calls are already wired to Search/Fetch, with optimizations also announced for Tencent CodeBuddy and WorkBuddy.
Path 2 — Claude Code or Codex with the DeepSeek backend. This is the most interesting case: these agents already hit the Search and Fetch APIs server-side, free of charge until November 1st. If you're using a terminal agent like DeepSeek TUI, check which search layer it uses — after November 1st, the difference will show up on your bill. Our selection of the best AI coding tools details the compatible harnesses.
Path 3 — Direct REST. For your own orchestrator, both endpoints are standard POSTs with Bearer authentication. Three tips:
- Start with
need_content: false. Thesnippetandsummaryare often enough, and fullcontentquickly bloats the context. - Use
rerank_scoreandauthorityfor local filtering, before any LLM call. - Keep
countlow (5-10) and let the agent refine by iterating — cheaper than acountof 50.
What about self-hosting in all this?
The weights are MIT, so the brain can run on your own hardware — antirez's DS4 inference engine makes DeepSeek V4 Flash usable locally. But the eyes remain with DeepSeek: the search engine is not published, and nothing suggests it ever will be. You can offload inference, not grounding. That's precisely where the dependency lies.
To build a complete agent with this kind of building block, our guide to the best autonomous AI agents compares the available architectures.
Common Mistakes
Mistake 1: Confusing Search and Fetch
Search discovers, Fetch extracts. Using Search to retrieve a URL you already know wastes calls and returns approximate results. Simple rule: known URL → Fetch; open-ended question → Search.
Mistake 2: Ignoring the 5 QPS cap
Launching ten agents in parallel that all search simultaneously is a guaranteed way to get error responses. Serialize searches at the orchestrator level or add a queue. The cap is per account, not per key or per agent.
Mistake 3: Building on free usage without planning for November 1st
Free access via Harness, Claude Code, and Codex ends on the evening of October 31, Beijing time. If your workflow runs on it, start logging your call volume now: it's your future cost line item. At $0.75 per thousand, the surprise will be a mild one, but it will still land.
Mistake 4: Swallowing the raw content field
A page's full content can amount to tens of thousands of tokens. On a 1M-context model, the temptation to pile everything in is strong — it's the surest way to turn $0.20 per task into $2. Filter by rerank_score first, and only fetch what survives.
Frequently Asked Questions
Do the Search and Fetch APIs work with any model?
Yes. These are self-contained REST endpoints with Bearer authentication: any orchestrator can call them, including with a model from another family. However, the free tier until November 1 only applies to calls routed server-side through DeepSeek Harness, Claude Code, or Codex.
Are failed requests billed?
No. The $0.75 rate applies per 1,000 successful requests. But the exact definition of "successful" deserves validation on your own use cases: measure your error rates before November 1 to anticipate your actual bill rather than the theoretical one.
Can I build an alternative search engine on top of it?
No, and that's not the goal. The cap of 5 requests/second per account and the proprietary index position Search/Fetch as a building block for agents, not as consumer-grade search infrastructure. For volume, you'll have to look elsewhere.
Do you need to switch to V4.1-Flash-1017 to use Search/Fetch?
No. Both APIs are model-independent. But the economic equation is designed to work as a whole: it's by combining abundant search with tokens at $0.15–0.30 per million that the cost per task stays around $0.20.
Do the open weights include the search engine?
No. Only the model weights are published under the MIT license on Hugging Face and ModelScope, as confirmed by Particle News. The search engine, its index, and its reranking remain proprietary, accessible only through the paid API.
Conclusion
By opening up the search layer its agents used internally at $0.75 per thousand calls, DeepSeek reminds us that the battle over AI agents is no longer just about the models, but about everything that powers them. Try Search/Fetch before November 1st: the free period is the best time to measure what your agent actually consumes. To keep up with what's changing each week, our roundup of AI news is made for exactly that.