研究一下,OpenAI最新推出的o3-deep-research(06-25),是否就是官网当中使用...

Erstellt am: 28. Juni 2025

Beantwortet mit O3 von Chat01

Frage

研究一下,OpenAI最新推出的o3-deep-research(06-25),是否就是官网当中使用o3的效果?

Gedanke

Crafting answer on o3-deep-research

I need to clarify a few things about o3-deep-research. It's a fine-tuned version of the o3 model, enhanced for agentic deep research tasks and tool interaction.

On the ChatGPT site, the base model is simply "OpenAI o3," but the deep research version shares the same core while adding specific tuning for tasks like browsing and Python processing.

For developers, there's also a model ID called "o3-deep-research" available since June 25.

This means that deep research might perform differently from o3, and it can be slower, with browsing integrated.

Crafting the response about o3-deep-research

o3-deep-research is a fine-tuned variant of o3, designed for deeper research tasks with integrated tool use. It's based on the same core weights but includes specialized tuning for browsing, Python, and data retrieval, perhaps offering longer context windows.

On the ChatGPT site, the "o3" references the base model, while "o3-deep-research" has been available for developers since June 25, 2025, offering autonomous search and citation capabilities. This version tends to perform slower than the base model due to its advanced features.

Thought for 53 seconds

Antwort

在 2025 年 6 月 25 日,OpenAI 在 API 与 ChatGPT 侧面向更多开发者和团队开放了 o3-deep-research 这一专用模型 ID,但它并不是一款“全新”模型;它是把官方网站所说的 o3 推理模型 进一步微调并加上工具链、长上下文与自主浏览能力后形成的“研究专用版本”。也就是说,两者共享同一主干权重与推理能力,只是在定位和默认工作流上有所差异。community.openai.com


o3-deep-research 的由来与发布时间线

日期里程碑说明
2025-02-02ChatGPT 内测上线 Deep Research作为按钮出现在 ChatGPT Pro 版,后台调用的是“即将发布的 o3 版本”openai.com
2025-02-12Altman 公开透露将下放 Plus 与免费额度每月 Plus 10 次 / 免费 2 次计划venturebeat.com
2025-04-24轻量版切换至 o4-mini 以扩大配额深度研究达到 25∼250 次/月,超过额度自动退化为轻量版openai.com
2025-06-25API 出现 o3-deep-research 模型名支持 Webhooks 与浏览工具链,定价与限额单独计算community.openai.com

技术关系:同一母体,不同“职业训练”

共同点

  • 核心模型:Deep Research 明确标注“基于 o3 推理模型微调”help.openai.com;官方博客也写道“Powered by a version of the upcoming o3 model”openai.com
  • 长思考与工具调用:o3 系列的卖点是“让模型在回答前思考更久并会用工具”openai.com,Deep Research 继承了这一能力并把“浏览器+Python”管线固化为默认步骤。reuters.com

区别

维度ChatGPT 里的「o3」Deep Research(o3-deep-research
交互方式即问即答,约 < 1 分钟返回接受任务→自主多轮浏览/代码→10–30 分钟生成报告theguardian.com
专用微调无(系统提示较轻)针对资料检索、整理、引用格式做强化学习aragonresearch.com
输出聊天回复,可调用工具带目录、引用、可能含表格/图像的长文报告writesonic.com
速率限制与普通聊天相同(依套餐)单独的月度 Task 配额,超额自动降级到轻量版openai.com

为什么 Deep Research“感觉更强”?

  1. 多步检索+归纳:它默认执行几十到上百次 web 搜索、读取 PDF/图表并用 Python 处理数据后,再开始写作,因而在资料完整性与可验证性上远超常规聊天reuters.com
  2. 显式透明的引用:Deep Research 报告每句话后都会附带来源 URL,方便人工复核,官方 FAQ 将这一点列为核心卖点help.openai.com
  3. 任务级记忆:流程化设计允许它在长达 30 分钟的会话中保持上下文与中间结论,比普通 o3 会话更不易“丢线索”。forum.openai.com

回答你的核心疑问

  • 是否“官网中的 o3 效果”?
    • 底层推理能力——是,同一代 o3 权重,数学、代码、视觉理解等基准分数几乎一致help.openai.com
    • 产品表现——不完全相同。Deep Research 通过更重的系统提示、浏览器与 Python 工具链,以及更长推理时间,换取了面向“研究”场景的深度与格式化输出。普通 o3 更偏实时、多模态对话与快速回答。

因此,你可以把 Deep Research 看成「o3 的研究员职业版」——基因相同,但职业训练与日常工作流不同,导致体验与定位都更像专业研究助手,而不是普通聊天机器人。

Frage

对比o3-deep-research API、o3 API、ChatGPT网页端o3、ChatGPT网页端deep research功能

Gedanke

Gathering deep research info

I need to find the latest details on deep research pricing and features, so I'll search for "o3-deep-research price API." Deep research relies on a combination of search and Python tools, which I’ll need to gather more about. I'll check the FAQ for relevant info, such as usage limits, and also review context window details (200k tokens). Finally, I'll pull at least 10 high-quality citations from diverse sources, such as openai.com, help.openai.com, and trusted news sites.

Organizing and gathering citations

I need to organize the ways I'll get citations. Here's my plan: I’ll gather sources like OpenAI’s community pages and pricing details, help pages on FAQs, and news sites like Axios, The Guardian, and Reuters. I'll also look for insights from blogs like MarketingAIInstitute and alexop.dev. With at least 15 diverse citations, I'll compile a summary that includes deployment channel, primary use case, pricing, and more. This summary will be clear and structured, using markdown with tables to clarify differences in API features like context window, output style, and query limits.

Thought for 57 seconds

Antwort

概览(一句话总结)
OpenAI 把同一颗 o3 推理模型打包成四种不同的“形态”:① 纯 API 模型 o3,② 配套的研究代理 o3-deep-research API,③ ChatGPT 网页端可选模型 o3,以及 ④ 网页端的“一键多步研究”按钮 Deep research。它们的底层权重一致,但在交付通道、默认工具链、速率/计费、任务时长与输出格式上做了差异化,以覆盖从秒级对话到半小时报告的不同场景。


1 模型与接入方式

形态入口 / Model ID默认流程典型时长适用场景
o3 APImodel="o3" via Chat Completions / Assistants单轮推理,可选 reasoning_effort,不自动浏览秒级代码、数学、聊天
o3-deep-research APImodel="o3-deep-research"自动多轮 Web 搜索 + 工具(浏览器、Python),支持 Webhooks & Background Mode5–30 min程序化生成长报告
ChatGPT o3ChatGPT 模型选择器对话式,多模态工具随用随调秒级高难即时问答、代码伴侣
ChatGPT Deep research“Deep research” 按钮先规划→循环检索→汇总成带引用的报告5–30 min竞品/学术/市场深研

community.openai.com help.openai.com help.openai.com


2 核心能力差异

2.1 工具链

  • o3 / ChatGPT o3 仅在被动调用时使用搜索、代码运行等工具,偏重快速回应。help.openai.com
  • Deep research 系列 内建“浏览器 + Python”流水线,循环检索、读 PDF/图表并分析后再写作;边栏实时展示思路与已访问网页。help.openai.com

2.2 输出

  • o3 返回普通聊天/JSON/函数调用结果。
  • Deep research 产生带目录、分节、脚注-URL-格式的长文报告,后续还会插入图像与图表。help.openai.com

3 上下文窗口与推理控制

  • 四者共享 200 k token 输入 / 100 k 输出 的超长上下文。help.openai.com
  • API 版本可通过 reasoning_effort(low/medium/high)调节思考步数;ChatGPT 界面隐式映射到“快速/标准/深思”。cookbook.openai.com

4 速率与使用额度

形态ChatGPT 额度¹API 速率²
ChatGPT o3100 条/周(Plus/Team/Enterprise)help.openai.com20 RPM / 200k TPM 起³
ChatGPT Deep researchFree 5 → 25 /月;Plus/Team/Enterprise 25 → 40 /月;Pro 250 /月(轻量版额外同量)help.openai.com
o3 API依层级给定 TPM/RPM,支持 Batch & Background
o3-deep-research API建议配合 Background Mode 避免 600 s 超时,支持任务 Webhook 回调。community.openai.com

¹ 额度会随扩容调整;企业订阅可单独谈判。
² 官方未公开详细表,以上为社区常见配额。
³ API Background Mode 可异步跑超长任务。community.openai.com


5 计费对比(2025-06-28)

形态计费方式最新公开价
o3 APItoken-based2/1Min2 / 1M in 8 / 1M out(官方页)openai.com <br> 0.40/0.40 / 1.60 新价已在论坛预告,尚未同步官网。community.openai.com
o3-deep-research API同 o3 token价;额外搜索内容 token 按输入价计费。openai.com
ChatGPT o3随订阅(Plus 20/mo;Pro20/mo; Pro 200/mo 等) —无额外计量。
ChatGPT Deep research初期 Pro $200/mo 含 100 次 — 现在随套餐免费,按上表任务限额,无额外费。theguardian.com openai.com

6 何时选哪一个?

目标推荐
秒级问答 / 对话编程 / 快速草稿ChatGPT o3 或 o3 API(低延迟、不必等待长流程)
自动化生成有来源的深度报告o3-deep-research API(可脚本化、批量跑、Webhook 回调)
零代码、一次性市场/论文/专利调研ChatGPT Deep research 按钮(UI 省心)
需要掌控提示、缓存、混编工具先用 o3 API,必要时自己 orchestrate 工具,而非 Deep research

7 最新动向与展望

  • Deep research API 于 2025-06-26 加入 webhooks 与官方 Agents SDK,方便把长时研究插入企业工作流。theverge.com
  • OpenAI 正在降低 o3 推理成本并推出高算力版 o3-pro,面向 30 分钟以上的重推理背景任务。community.openai.com
  • ChatGPT 端已上线 轻量版 deep research(o4-mini 驱动),超过额度后自动降级,兼顾更多免费与 Plus 用户。theverge.com

参考文献

(已在文中逐句标注)


小贴士:如果你只是想把 Deep research 当作“API 里的网搜助手”,记得在请求体里加入 tools=[{"type":"web_search"}],并为长任务启用 background=true,再配合 Webhook 获取最终报告,以免超时。

Frage

三个个问题:
第一,是否可以说,o3的API与ChatGPT官网的o3的区别在于,后者是多轮思考以及可以使用工具。
第二,如果是,那么新的o3-deep-research API是否也可以多轮思考以及可以使用工具。
第三,如果是,o3-deep-research API它使用的工具是openai后台提供的,还是要用户提供?

Gedanke

Clarifying differences between o3 API and ChatGPT o3

The main difference between the o3 API and ChatGPT o3 is not about capabilities but about the interface and tool availability. Both can handle multi-step reasoning, and the API can also use tools, if configured. The key distinction is that ChatGPT UI provides built-in orchestration for multi-turn thinking and automatic tool usage (like browsing or code), whereas the o3 API requires extra setup for these tools. The default environment influences user convenience more than the underlying abilities.

Summarizing response on o3-deep-research API and tools

To clarify, the o3-deep-research API does support multi-step reasoning and can use tools, particularly for in-depth research tasks. It autonomously searches and reads, which is a specialized feature for deep research. The tools (like browsing and Python) are hosted by OpenAI, so the user does not need to implement them, though they can add custom tools as needed. Numerous citations reinforce these features, including official OpenAI documentation, tutorials, and community posts.

Thought for 45 seconds

Antwort

简要结论

  • 不是:把“ChatGPT 网页端 o3”与“o3 API”简单地说成“前者能多轮思考和用工具而后者不能”并不准确——两者同样具备长链式推理(reasoning_effort 参数)与工具调用能力,只是 ChatGPT UI 默认把这些能力包装成一键可用,而 API 端需要开发者在请求里显式开启或声明工具。
  • o3-deep-research API 专门为“多轮检索+分析+写报告”场景定制,天然会在一次调用里自行循环调用浏览器、Python 等托管工具
  • 托管 vs 自定义o3-deep-research 默认使用 OpenAI 后台托管web_searchpython 等工具;开发者若有需要,也可以像普通 o3 那样再挂自定义 Function/MCP 工具。

1. o3 API vs ChatGPT 网页端 o3

维度o3 APIChatGPT o3(网页端)
内核模型同一套 o3 权重同左
多轮推理通过 reasoning_effort (low/medium/high) 控制,支持在一次调用里花更多“隐藏推理 token”ai-sdk.dev devblogs.microsoft.comUI 隐式选择推理档位;用户无需写参数openai.com
工具调用可调用托管或自定义工具;需在请求 JSON 里列出 toolsallowed_toolsopenai.com openai.comChatGPT 已预接入浏览器、代码解释器、DALL·E 等工具,模型自动发现并组合使用openai.com
开发者工作量需自己编排对话轮次、工具列表、长任务(可用 Background Mode)openai.comUI 自动管理上下文与工具,零代码

因此,“多轮思考+用工具”不是 ChatGPT o3 的独占特性,而是 API 端可选、UI 端默认开启的体验差异。DataCamp 的实战教程也展示了在 API 中让 o3 连续调用搜索与 Python 解决复杂题的流程datacamp.com


2. o3-deep-research API 是否具备多轮思考与工具调用?

是的。官方 FAQ 说明,Deep Research 基于 o3 微调,能自主在公共网页检索—阅读—汇总,输出带引用的长报告help.openai.com
OpenAI 开发者社区建议在调用时配合 Background Mode 与 Webhooks,因为一次任务可能持续 5–30 分钟并触发大量工具调用community.openai.com。新闻报道也把它描述为“几分钟内完成人类数小时研究的链式流程”reuters.com,实际体验等待 5–30 分钟可得高质量报告businessinsider.com


3. o3-deep-research API 使用的工具来源

类型提供方说明
托管工具OpenAIAgents SDK 文档列出 web search、retrieval、computer use、code interpreter 等“Hosted tools”在 OpenAI 服务器侧运行,无须开发者部署openai.github.io openai.github.io。Deep Research 默认启用其中的 浏览器 + Python 流水线help.openai.com
自定义 Function/MCP 工具开发者可选Responses API 允许把本地函数或远程 MCP 服务挂载为工具,再通过 allowed_tools 让模型调用cookbook.openai.com openai.com。这与普通 o3 API 的自定义流程一致。

换言之,开箱即用的就是 OpenAI 托管工具;若业务还需要访问私有数据库、第三方 API 等,可按常规方式把自家函数注册进去,与 Deep Research 的浏览器/Python 并存。


参考来源

  1. OpenAI 官方《Introducing o3 and o4-mini》openai.com
  2. Deep Research FAQ, OpenAI Help Centerhelp.openai.com
  3. OpenAI 社区帖《Deep research in the API, webhooks and web search with o3》community.openai.com
  4. OpenAI 官网《New tools & features in the Responses API》openai.com
  5. Agents SDK Docs – Hosted tools sectionopenai.github.io openai.github.io
  6. DataCamp《OpenAI o3 API: Step-by-Step Tutorial》datacamp.com
  7. OpenAI Cookbook《Better performance from reasoning models using the Responses API》cookbook.openai.com
  8. Reuters《OpenAI launches new AI tool to facilitate research tasks》reuters.com
  9. Business Insider《I tried ChatGPT’s new Deep Research…》businessinsider.com
  10. AI-SDK.dev 指南《Refining reasoningEffort for o3-mini》ai-sdk.dev

Diesen Q&A teilen