[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-search-fetch-api-075":3,"topics-all":38,"news-related-6678ad05-308c-4440-beac-a1571f4c8059":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"6678ad05-308c-4440-beac-a1571f4c8059","DeepSeek 把搜索和抓取 API 开放:每千次 0.75 美元,Agent 工具栈拼图补齐","DeepSeek 把自研搜索引擎拆成 Search 和 Fetch 两个 API 对外开放,每千次成功调用 0.75 美元,失败不计费。通过 DeepSeek Harness、Claude Code、Codex 走 server-side 调用的免费窗口到 10 月 31 日截止。","10 月 17 日 DeepSeek 在 V4.1-Flash-1017 模型升级之外,同步上线了 Search 和 Fetch 两个独立 API——把每天承接数亿请求的自研搜索引擎,拆成两条 JSON 接口卖给第三方,每千次成功调用 0.75 美元,失败不计费。\n\n## 接口分工清晰\n\nSearch 和 Fetch 共用同一个 DeepSeek API Key,走 Bearer 鉴权,base URL 也是 https:\u002F\u002Fapi.deepseek.com,Search 命中 `\u002Fv1\u002Ftools\u002Fsearch`,Fetch 命中 `\u002Fv1\u002Ftools\u002Ffetch`。两条路径独立计费、独立 QPS 限额,互不挤兑。\n\nSearch 接受 1 到 128 字符的 query,默认返回 10 条、最多 50 条,每条带 snippet(100 到 200 字符)、summary(600 到 800 字符),`need_content=true` 时再带回整页 content。它还支持完整搜索语法:`site:`、`-site:`、`lang:`、`after:`、`before:`,双引号精确匹配,熟悉 SEO 的人零迁移成本。\n\nFetch 更直白:扔一个 URL 进去,返回 `{title, content}`,content 直接是 Markdown——专门为模型偏好 Markdown 设计。错误码语义跟 HTTP 不一致:404 不一定代表页面消失,也可能只是读不到内容;403 是被安全策略拒掉;402 才表示欠费。Agent wrapper 要把 4xx 单独处理,别让 agent 自己脑补\"网页没了\"。\n\n## 价格、限额和那扇免费窗口\n\n每千次成功请求 0.75 美元,失败、空结果、超时都不计费——这一条要圈出来。Search QPS 上限 100,Fetch QPS 上限 25,与模型并发分开池计费,不再撞模型那一边。\n\n免费窗口更狠。10 月 17 日到 10 月 31 日 16:00 UTC 之间,凡是走 DeepSeek Harness、Claude Code、Codex 的 server-side 工具调用,Search\u002FFetch 都不计费。DeepSeek 借评测渠道把搜索通路塞进 Claude Code、Codex、腾讯 CodeBuddy、WorkBuddy 这些对手的编程 agent 里。31 日一过,起步价一刀切到 0.75 美元\u002F千次。\n\n同日 Startup Fortune 列出的市场价:Brave 大约 5 美元\u002F千次,Exa 4 到 7,Perplexity 1 到 5,Tavily 8 到 16。即便只看下限,DeepSeek 也便宜一个数量级。不过首发价后续可能调,DeepSeek 早前在别的产品线上调过价,开发者要心里有数。\n\n## 真正值得警惕的是数据流向\n\n便宜到 0.75 美元\u002F千次是有代价的。DeepSeek 隐私政策写得很直白:收集、处理、存储个人数据都在中国境内,平台条款适用中国法律。API 用户输入是否被用作训练,API 条款里没写清楚;Search 指南也未提留存期和区域限制。意大利、德国此前已就数据出境对 DeepSeek 采取过行动——查询里一旦带客户数据、机密信息、未公开数据,先读条款再调用。\n\n技术侧是干净的:Bearer 鉴权、OpenAI 兼容错误格式、JSON in \u002F JSON out。模型侧 DeepSWE v1.1 上 77.0 比 Anthropic Opus 5 的 74.0、74.2% Pass@1 比 Opus-5 的 74.0%——这些都是厂商自报数,独立评测还没跟上。\n\n这次发布的信号不是\"DeepSeek 又升级模型了\",而是\"DeepSeek 在做底层工具栈\"。Search\u002FFetch 加 Harness 加 V4.1-Flash-1017,串起来是一套完整的低成本 agent 拼图。Agent 一侧真正要算的账还是 QPS 能不能撑住实际流量、敏感数据要不要走这条路,以及 31 日之后免费窗口关闭,这张牌会不会变。","https:\u002F\u002Fapi-docs.deepseek.com\u002Fguides\u002Fsearch\u002F","3202bafd-f422-4632-b981-ec60896b79df",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"731a61f4-57ae-4008-872d-756297850abe","en","DeepSeek opens Search and Fetch APIs at $0.75 per 1,000 calls, free for Claude Code and Codex through October 31","DeepSeek broke its in-house search engine out as two standalone APIs—Search and Fetch—at $0.75 per 1,000 successful calls, failed requests not billed. Server-side calls routed through DeepSeek Harness, Claude Code, and Codex stay free until 16:00 UTC on October 31.","On October 17, 2026, alongside the V4.1-Flash-1017 model upgrade, DeepSeek launched two standalone APIs—Search and Fetch—breaking out an in-house search engine that already handles hundreds of millions of requests per day into two JSON endpoints priced at $0.75 per 1,000 successful calls, with failed requests billed nothing.\n\n## Two endpoints, clean separation\n\nSearch and Fetch share the same DeepSeek API Key via Bearer auth, both rooted at https:\u002F\u002Fapi.deepseek.com. Search hits `\u002Fv1\u002Ftools\u002Fsearch`, Fetch hits `\u002Fv1\u002Ftools\u002Ffetch`. Each has its own billing, its own QPS cap, and the two quotas do not interfere.\n\nSearch accepts a 1 to 128-character query and returns 10 results by default, capped at 50. Each item ships three layers: a UI-oriented snippet (100 to 200 characters), a model-oriented summary (600 to 800 characters), and full-page content when `need_content=true` is set. The endpoint also supports a familiar search operator grammar: `site:`, `-site:`, `lang:`, `after:`, `before:`, exact phrases, and double-quoted exclusions—anyone who has done SEO work picks it up with zero migration.\n\nFetch is even simpler: hand it a URL, get back `{title, content}` where content is already Markdown, not HTML, because models prefer Markdown. The HTTP error semantics here are subtly different from the standard: a 404 does not necessarily mean the page is gone, only that the content could not be extracted; 403 means the security policy blocked the request; 402 means the account is out of credit. Agent wrappers need to handle 4xx codes as retrieval issues, not as content disappearance, or the agent will draw the wrong conclusion.\n\n## Pricing, limits, and a free window\n\n$0.75 per 1,000 successful requests, with failed requests, empty results, and timeouts not billed—that line is the takeaway. Search is capped at 100 QPS, Fetch at 25 QPS, and these pools are separate from model concurrency, so agent fan-outs stop clobbering model rate limits.\n\nThe free window is even more aggressive. Between October 17 and 16:00 UTC on October 31, any server-side tool call to Search or Fetch made through DeepSeek Harness, Claude Code, or Codex is not charged. DeepSeek is using this window to push its search pipeline into Claude Code, Codex, Tencent CodeBuddy, and WorkBuddy—other companies' coding agents. After 16:00 UTC on October 31, the list price snaps back to $0.75 per 1,000 calls.\n\nFor context, Startup Fortune's launch-day comparison lists Brave around $5 per 1,000 calls, Exa $4 to $7, Perplexity $1 to $5, and Tavily $8 to $16. Even by the lower end of those ranges, DeepSeek is roughly an order of magnitude cheaper. Launch prices are not promises, however—DeepSeek has raised prices on other product lines before, and developers should keep that in mind.\n\n## The data question is the real one\n\nThere is a trade-off for that $0.75 price. DeepSeek's privacy policy is explicit that personal data is collected, processed, and stored in mainland China, and the platform terms are governed by Chinese law. Whether API inputs feed model improvement is not stated clearly in the API terms; the Search guide says nothing about retention windows or data residency. Italian and German regulators have previously acted against DeepSeek over data transfers to China—if your queries carry customer data, confidential information, or anything not yet public, read the terms before you wire this up.\n\nThe technical surface is clean: Bearer auth, OpenAI-compatible error format, JSON in, JSON out. The model-side claims, like a 77.0 DeepSWE v1.1 score beating Anthropic Opus 5 at 74.0, or a 74.2% Pass@1 on DeepSWE v1.1 edging Opus-5's 74.0%, are all vendor-reported numbers on tables that leave out newer competitors—independent testing has not yet caught up.\n\nThe signal here is not \"DeepSeek upgraded a model again.\" It is \"DeepSeek is building a tool layer.\" Search, Fetch, Harness, and V4.1-Flash-1017 stitch together into a complete low-cost agent puzzle. The question for anyone building agents is whether their actual workloads fit under the limits, whether their sensitive data should travel through this route at all, and whether this price card will still be in play once the October 31 free window closes.","deepseek-search-fetch-api-075","2026-10-20T08:00:00Z","2026-10-20T01:16:17.544940Z","2026-10-20T01:16:17.544947Z",true,"agent",174,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"9fcb5d74-315e-4d57-a7da-290fbfb9e91c","阿里 Qwen 发布弹性 RL 框架 QwenGyre,训练长时 Agent 提速 1.85x","alibaba-qwen-qwengyre-elastic-rl-xlong-agents","2026-10-14T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"5edbfca7-7c1a-4df3-8577-a0195408b827","RSI-Anything:世界模型陪跑,自我改进有了回滚键","rsi-anything-world-model-recursive-self-improvement","2026-10-12T21:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b95bf7cf-e962-49e8-847c-2e45c917f19d","S-Space:多模态模型看得懂空间,却转不过弯","s-space-multimodal-spatial-workspace","2026-10-12T17:19:44+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"fc533af5-9e8a-43c2-8408-2f931a6aa398","EngramEdit:把事实塞进 LLM 的记忆抽屉","engramedit-decoupled-knowledge-llm","2026-10-09T04:00:00+00:00"]