OpenAI 推出 Deep Research

2025 年 2 月，OpenAI 发布了 Deep Research。这是一个专门干「调研」的 Agent：你抛出一个研究问题，它会自己规划搜索路径、打开网页、阅读资料、反复迭代，最后生成一份带引用的研究报告。整个过程不需要你操作浏览器，你只需要等待——从几分钟到几十分钟，一份「像研究员写出来」的文档就摆在你面前。

在 Deep Research 出现之前，「让 AI 帮忙查资料」的体验是碎片化的：你得不断提问、复制粘贴、自己整理。Deep Research 把整条流水线自动化了——搜索、阅读、判断、整合、写作，串成一个端到端的自主流程。它基于 o3 系列推理模型构建，每次任务会运行很长时间，用「多轮思考 + 多轮检索」来逼近一份像样的调研报告。

它发布时最让人感叹的是「持久性」。过去的 AI 回答是「一次性」的：问完就结束。Deep Research 则像一位能连续工作很久的研究助理，任务越复杂，它思考越久。这种「长任务 Agent」的形态，和 Operator 的浏览器操作、Codex 的写代码一样，都指向同一个方向：AI 正在从「回答者」变成「执行者」。

当然，Deep Research 不是没有争议。它会幻觉、可能误读来源、报告质量参差不齐；而且一次任务消耗的算力巨大，初期只面向最高档付费用户。但作为一个产品形态，它验证了一件事：深度调研是可以被 AI 产品化的场景，而且用户愿意为此付费。Google、Perplexity 随后纷纷推出类似功能，研究型 Agent 成了 2025 年最热的产品赛道之一。

回看 Deep Research 的发布，它的价值在于定义了一种新的产品范式：AI 不只是更快地回答问题，而是能替人完成一整段耗时费力的智力工作。当调研、写报告这些过去要花几小时的事被压缩成「等一个 Agent 跑完」，信息工作的形态，正在被悄悄改写。

In February 2025 OpenAI released Deep Research, an agent dedicated to research. You pose a question, and it plans its own search paths, opens web pages, reads material, iterates, and finally produces a cited research report. You never touch the browser—you just wait, from minutes to tens of minutes, and a document "written like a researcher" appears before you.

Before Deep Research, "let AI help me research" was a fragmented experience: endless prompts, copy-paste, manual organizing. Deep Research automated the entire pipeline—search, reading, judgment, synthesis, writing—into an end-to-end autonomous flow. Built on the o3 reasoning family, each task runs for a long time, using multi-step thinking and multi-step retrieval to approximate a proper research report.

The most striking thing at launch was persistence. Previous AI answers were one-shot: ask, done. Deep Research was like an assistant that keeps working for a long time—the more complex the task, the longer it thinks. This long-horizon agent form, alongside Operator's browser operation and Codex's coding, all pointed the same way: AI was moving from "answerer" to "executor".

Deep Research was not without controversy. It hallucinates, may misread sources, and report quality varies; each task also burns enormous compute, and it initially launched only for top-tier subscribers. But as a product form it proved something: deep research is a scenario AI can productize, and users will pay for it. Google and Perplexity quickly shipped similar features, making the research agent one of 2025's hottest product categories.

Looking back at Deep Research, its value was defining a new product paradigm: AI doesn't just answer questions faster—it can complete a whole stretch of time-consuming intellectual work on your behalf. When research and report-writing tasks that used to take hours get compressed into "wait for an agent to finish", the shape of knowledge work is being quietly rewritten.