跳到主要內容

DeepSeek搜尋API評測:千次0.75美元,比Tavily便宜10倍 | DeepSeek Search API Review: $0.75/1K vs Tavily's $8

By Kit 小克 | AI Tool Observer | 2026-10-19

🇹🇼 DeepSeek搜尋API評測:千次0.75美元,比Tavily便宜10倍

DeepSeek搜尋API在2026年10月17日悄悄上線,把內部搜尋引擎拆成兩個獨立端點:Search(回傳排序後的網頁結果)與Fetch(把整頁轉成Markdown),單次成功呼叫只收0.75美元/千次,比業界常見的Tavily(8美元/千次)便宜了近10倍。對需要即時上網查資料的AI代理(agent)和RAG系統來說,這是目前市面上最便宜的選項之一——但便宜背後藏著幾個先搞懂比較保險的限制。

DeepSeek搜尋API多便宜?定價怎麼算?

Search和Fetch各自收費0.75美元/千次成功呼叫,只有在回傳HTTP 200且內容非空時才計費,逾時、失敗或空結果都不收錢。部分開發工具透過整合管道使用可免費用到2026年11月1日(日本時間凌晨1點),之後才正式開始計費。相比Tavily每千次8美元的定價,DeepSeek等於用十分之一的成本提供網路存取能力。

速率限制會不會卡住你的代理?

每個帳號硬性限制5 QPS(每秒5次請求),即使開多組API金鑰也無法疊加上限,換算下來每天最多43.2萬次呼叫。這個量級適合個人應用或單一coding agent,但不適合拿來做大規模爬蟲或多租戶服務——如果系統需要同時服務多個客戶,得自己做排隊或分流。另外query與ori_question欄位都被限制在128個Unicode字元內,代理如果習慣生成長查詢字串,沒先截斷就會收到400錯誤。

搜尋品質真的夠用嗎?

這是整件事最容易被忽略的細節:計費邏輯只看「有沒有回傳內容」,不看內容「有沒有用」。也就是說,就算搜尋結果是十筆完全不相關的網頁,只要非空就照樣算成功、照樣收費。有開發者實測,DeepSeek搜尋與Exa相比,在29個查詢、12個平台的比對中重疊率只有2.1%——對中文來源抓取快、夠新,但YouTube、WeChat等平台覆蓋明顯偏弱。建議接上rerank_score做二次過濾,別讓低相關度的搜尋結果直接塞進模型的context window,否則省下的API費用,很可能都賠在多花的token和重試成本上。

資料隱私與合規怎麼看?

DeepSeek的伺服器設於中國,代表每一次搜尋關鍵字與抓取的網址內容都會經過當地伺服器處理。如果應用是面向歐美客戶的商用產品,這一點在合規審查時幾乎必然會被問到,上線前最好先問過法務或資安團隊,而不是等上線後才補救。

該不該換成DeepSeek搜尋API?

如果是個人專案、內部工具或流量不大的coding agent,DeepSeek搜尋API的價格確實誘人,值得拿來做A/B測試。但如果系統要服務企業客戶或有合規要求,搜尋品質與資料主權的落差,可能比每千次省下的7塊多美金更重要。好不好用,試了才知道——建議先用自己實際會問的問題做小規模測試,再決定要不要整組換掉現有的搜尋服務。


🇺🇸 DeepSeek Search API Review: $0.75/1K vs Tavily's $8

DeepSeek quietly opened its in-house search engine to developers on October 17, 2026, splitting it into two standalone endpoints: Search (ranked web results) and Fetch (turns a page into Markdown). Each costs just $0.75 per 1,000 successful calls — about 10x cheaper than Tavily's $8 per 1,000. For AI agents and RAG pipelines that need live web access, this is one of the cheapest options on the market right now. But the price tag hides a few catches worth knowing before you switch.

How Much Does the DeepSeek Search API Actually Cost?

Search and Fetch are each billed at $0.75 per 1,000 successful requests, charged only when the call returns HTTP 200 with non-empty content — timeouts, errors, and empty results are free. Some integrations route through third-party dev tools for free until November 1, 2026 (1am Japan time), after which standard billing kicks in. That's roughly a tenth of what Tavily charges for a comparable web-access API.

Will the Rate Limits Choke Your Agent?

Each account is hard-capped at 5 requests per second, and stacking multiple API keys doesn't raise that ceiling — the practical maximum is about 432,000 calls per day. That's fine for a personal project or a single coding agent, but not built for bulk crawling or multi-tenant workloads. On top of that, the query and ori_question fields are capped at 128 Unicode characters, so agents that generate long search queries will hit 400 errors unless you truncate first.

Is the Search Quality Actually Good Enough?

This is the part most headlines skip: billing only checks whether a response is non-empty, not whether it's useful. Ten irrelevant results still count as a successful, billable call. One developer's side-by-side test found only 2.1% overlap with Exa across 29 queries on 12 platforms — fast and fresh for Chinese-language sources, but noticeably weaker on YouTube and WeChat. The practical fix is filtering results through rerank_score before they reach your model's context window; otherwise the tokens and retries burned on junk results can erase the savings from the cheaper API calls.

What About Data Privacy?

DeepSeek's servers are based in China, meaning every search query and fetched URL is processed there. If you're shipping a customer-facing product for US or EU users, that's a compliance question your legal or security team will almost certainly raise — better to ask before launch than after.

Should You Switch to the DeepSeek Search API?

If you're running a side project, internal tool, or a low-traffic coding agent, the price is genuinely worth an A/B test. But if you're serving enterprise customers or operating under data-residency rules, the gap in search quality and data sovereignty may matter more than the $7.25 you save per thousand calls. As always: you won't know until you try it — test it against the actual questions your app will ask before ripping out your current search provider.

Sources / 資料來源

常見問題 FAQ

DeepSeek搜尋API多少錢?

Search和Fetch各0.75美元/千次成功呼叫,只在回傳非空內容時才計費,比Tavily每千次8美元便宜約10倍。

DeepSeek搜尋API有速率限制嗎?

有,每個帳號上限5 QPS(每天最多43.2萬次),多組API金鑰無法疊加上限,查詢字串也限制在128字元內。

DeepSeek搜尋API的搜尋品質如何?

實測顯示與Exa重疊率僅2.1%,中文來源快且新,但YouTube、WeChat等平台覆蓋較弱,建議搭配rerank_score過濾再使用。

用DeepSeek搜尋API有資料隱私風險嗎?

有,伺服器設於中國,查詢內容與抓取頁面都會經當地伺服器處理,面向歐美客戶的商用產品上線前建議先做合規審查。

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code