10 月 30 日，百度 APP 围绕搜索框、AI 镜头、GUI 可视化、AI 互动应用等方向推出多项升级，并全新升级搜索 AI API。价格方面，搜索 AI API 增强版接口活动期间最低为 4.9 元/千次，旗舰版为 9.9 元/千次。百度称，4.9 元/千次的价格低于 DeepSeek 搜索 API，希望通过更低调用成本扩大搜索能力在 Agent 中的使用。
与此同时，百度还进一步开放 API、MCP、Skill 等多种接入方式，希望把搜索能力从百度 APP 内部释放给更多 Agent 和企业应用。百度表示，搜索正在从“一次提问、一次回答”，进一步走向“一项任务、一份交付”，同时通过多 Agent 协同和在线反馈闭环实现递归式自我进化。
搜索 AI API 升级：4.9 元/千次，搜索开始 RSI
面向开发者和企业，百度全新升级的搜索 AI API，其背后是一套 AI 原生四层架构。
在内容层，百度汇聚全球 80 余种语言、千亿级优质索引，并结合大模型进行内容判断，官方数据显示数据可信率提升超过 50%；在引擎层，百度重点加入了在线反馈闭环和多 Agent 协同自进化机制。每一次真实检索以及随后的用户反馈，都可以沉淀为新的训练信号，再通过“检索—反馈—再训练”的闭环重新回流模型，让系统持续调整检索和结果生成策略。
百度将这一过程称为“RSI”：搜索系统不仅响应用户问题，同时还利用真实使用数据不断产生新的训练信号，在无需人工逐条干预的情况下持续优化。官方数据显示，这套机制使整体迭代效率提升 80%以上。
在架构层，搜索 AI API 强调高并发、低时延和企业级服务保障，支持百万级并发调用；在应用层，则进一步开放 API、MCP、Skill 等多种接入方式，方便 Agent 和企业系统调用。百度披露，文心助手接入相关能力后，整体效果提升 20%。
百度搜索 AI API 将增强版价格压到 4.9 元/千次后，很快引发社区讨论。多数人认为，真正值得关注的并不是每千次便宜 0.1 元，而是百度和 DeepSeek 开始争夺 Agent 背后的默认搜索入口。
有人指出，百度 4.9 元目前属于限时活动价，不能直接等同长期价格。此外，对开发者而言，一毛钱的价差也不足以决定迁移，真正影响选择的是搜索结果质量、权威性、时效性、调用稳定性，以及完成一个真实 Agent 任务究竟需要多少次检索。
搜索不只给答案，开始直接交付结果
百度认为，搜索正在经历从“找信息”到“办事情”的变化。
过去，用户面对一个复杂问题，往往需要多次搜索、打开多个网页、手动比较信息，再自己完成后续操作。现在，百度希望把这一流程压缩成更短的链路：复杂问题交给搜索进行多步查找和整理，直接返回经过核实、汇总后的结论；高频事务则直接交付结果；需要长期关注的事项，则可以在用户确认后持续跟踪，在发生变化时主动提醒。
百度将这一能力概括为三件事：复杂的事“替你查到底”、惦记的事“替你持续盯”、要做的事“直接交成果”。目前，这类能力已经覆盖 170 余种工具场景。
这种设计也体现出百度对 Personal Agent 边界的一种判断：Agent 可以负责持续检索、整理信息和准备方案，但在支付、取消订单、改变行程等关键节点，最终决定权仍然留给用户。
百度在会上将搜索的发展划分为三个阶段：最初解决“一次提问”，随后开始完成“一项任务”，而现在则继续走向 Personal Agent。百度认为，Personal Agent 真正的价值在于记住用户背景、理解长期需求，并主动关注用户在意的变化。
除了搜索框，百度还重点升级了 AI 镜头。
百度将传统拍照识物的问题归结为三个“断点”：看不全、说不清、问不深。拍照只能捕捉一个瞬间，但现实世界始终在运动；用户最想问的对象往往又恰恰是不知道名字的东西；即使识别出了名称，也很难围绕它继续深入提问。
针对这些问题，升级后的百度 AI 镜头开始从“定格”走向“流动”。首先是多目标同步识别和动态追踪。比如动物园里跳动的猴子、海洋馆中游动的鱼群，也可以在移动过程中持续保持标签跟随，用户不需要再逐个拍照。其次是交互方式发生变化。用户可以直接点一下、圈一下，再围绕目标提问。镜头还能够在拍照之后无缝切换到实时对话，继续追问。
百度此次还进一步强化了 AI 镜头的语音交互能力，支持东北话、四川话、粤语、长沙话等 9 种方言进行实时对话。
GUI 覆盖提升 5 倍
这次升级的另一个重点是搜索结果页的 GUI 化。百度表示，搜索页 GUI 场景覆盖提升 5 倍，已经接入政务、民生、央行等超过 50 亿条权威资源数据，覆盖 20 多个行业。
传统搜索结果往往以文本和链接为主，但在金融、政策、数据分析、公共服务等场景中，用户真正需要的并不只是找到一段文字，而是快速看懂数据之间的关系。因此，百度尝试利用 AI 智能可视化和专业图表模板，将复杂信息转换为图表、卡片和可交互界面，让数据“看得见、能比较、可交互”。
这也意味着搜索结果开始从“内容页面”逐渐变成“应用页面”。用户不用先跳转到其他网站或工具，再完成计算、筛选和比较，越来越多操作可以直接在结果页完成。
百度表示，这套 GUI 能力将继续向生态开放，权威数据机构和专业合作伙伴都可以接入。
把 AI 互动内容塞进信息流
内容侧也是这次百度 APP 升级的重要方向。百度正式推出 AI 互动剧和 AI 互动应用，希望把内容从单向消费变成用户参与。
AI 互动剧允许用户在关键剧情节点选择不同分支，从而改变后续故事走向。比如用户面对故事中的关键选择，可以自己决定，不同选择会进入完全不同的剧情。AI 互动应用则更接近轻量化应用和小游戏，可以把文字、角色、知识和玩法组合起来。
百度希望把这类互动内容直接接入信息流推荐。搜索负责理解用户主动提出的问题，推荐则判断用户在浏览过程中刚刚产生的兴趣，再把相关互动剧和互动应用推给用户。背后依托百度信息流的多天 Agent 推荐系统，希望把“我也想试试”的瞬间兴趣直接转化为下一次参与。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On October 30, Baidu APP launched several upgrades around the search box, AI shot, GUI visualization, AI interactive application, and completely upgraded the search for AI API. In terms of price, the minimum duration of the search for the AI API enhanced interface activity was 4.9 yuan per thousand and the flagship version 9.9 yuan per thousand. Baidu states that the price of 4.9 yuan per thousand is lower than DeepSeek ' s search API and that it is hoped that the use of search capabilities in Agent will be expanded through lower call costs.
At the same time, Baidu has further opened up various modes of access, such as API, MCP, Skill, in the hope of releasing search capabilities from within Baidu APP for more Agent and business applications. Baidu states that the search is moving from “one question, one answer” to “one mission, one delivery” and is evolving incrementally through multiple Agent collaborative and online feedback loops.
Search AI API Upgrade: 4.9 yuan/thousands, search start RSI
For developers and businesses, Baidu is a new upgrade of the search AI API, behind which is a set of four-storey structures.
At the content level, Baidu brings together over 80 languages, hundreds of billions of graded high-quality indexes worldwide and makes content judgements in conjunction with LLM, where official data show an increase of over 50 per cent in data credibility; at the engine level, Baidu focuses on online feedback closed loops and multiple Agent collaborative self-evolution mechanisms. Each real search and subsequent user feedback can sink into a new training signal and allow the system to continuously adjust its retrieval and result generation strategies through the "Retrieving - Feedback - Retrain" closed loop reflow model.
Baidu refers to the process as “RSI”: not only does the search system respond to user problems, it also produces new training signals using real-life data and continuously optimizes without the need for manual article-by-article intervention. Official data show that this mechanism has resulted in more than 80 per cent improvement in overall iterative efficiency.
At the architecture level, the search for AI API emphasizes high-symmetry, low-latency and enterprise-level service security to support million-scale and call-up; at the application level, multiple access methods, such as API, MCP, Skyll, are further open to Agent and enterprise systems. Baidu reveals that the overall effect of Ernie's assistant's ability to access it has increased by 20%.
Baidu search AI API quickly triggered community discussion after Baidu pushed the enhanced version to 4.9 yuan/thousands of times. Most people believe that what really deserves attention is not a zero per thousand, but Baidu and DeepSeek are fighting for the default search portal behind Agent.
It was noted that Baidu 4.9 was currently the activity price of limited-time and could not be directly equivalent to the long-term price. Moreover, for developers, the price differential of a dime is not sufficient to determine migration, and what really affects the choice of search results is quality, authority, timeliness, call stability, and how many searches are required to complete a true Agent mission.
Search not only answers, but direct delivery.
Baidu believes that the search is going through a change from “seeing information” to “doing things”.
In the past, users faced a complex problem, often requiring multiple searches, multiple web pages, manual comparison of information and subsequent follow-up on their own. Now, Baidu wants to compress the process into a shorter chain: complex issues are referred to search for multiple steps to search and sort, and direct back to verified and aggregated conclusions; HF services deliver results directly; matters requiring long-term attention can be tracked on a continuous basis after confirmation by the user and be proactively alerted to changes.
Baidu summed up this ability as three things: complex things: “to find out for you”, thoughts about “to keep your eyes on” and what to do “to deliver directly”. Currently, these capabilities cover more than 170 tool scenarios.
This design also reflects the judgement of Baidu on the personal Agent boundary: Agent can be responsible for the continuous retrieval, collation and preparation of information and programs, but the final decision remains with the user at key points such as payment, cancellation of orders, change of itinerary, etc.
At the meeting, Baidu divided the development of the search into three phases: an initial “question-and-question” resolution, followed by a “mission-to-task” and now continuing to move towards the personal Agent. According to Baidu, the real value of Personal Agent lies in remembering the background of the user, understanding the long-term needs and proactively following the changes that the user is concerned about.
In addition to the search box, Baidu focused on upgrading AI shot.
Baidu lumps the problem of traditional photographic objects into three “breakpoints”: incompleteness, confusion and inexplicability. Photo captures only one moment, but the real world is always moving; the people most interested in asking are often precisely those who do not know the name; even if the name is identified, it is difficult to continue asking questions about it.
In response to these problems, the upgraded Baidu AI Z87Z began to move from a “modified” to a “mobile”. The first is multi-target simultaneous identification and dynamic tracking. For example, in a zoo, monkeys that jump, fish that move in a seahouse can also keep their tags on a continuous basis and users do not need to take photographs one by one. The second is an interactive change. Users can click, circle and ask questions around the target. shot was also able to sew it to real-time conversations after taking pictures and continue to ask questions.
Baidu further enhanced AI shot's voice-interaction capability, supporting real-time dialogue among nine dialects, including North-East, Sichuan, Sichuan, Mandarin and others.
GUI Overwrite 5 times higher
Another focus of this upgrade is the GUIization of the search results page. According to Baidu, the search page GUI coverage has increased fivefold, with access to more than 5 billion authoritative resource data from government, people's lives, central banks, etc., covering more than 20 industries.
Traditional search results are often dominated by text and links, but in such contexts as finance, policy, data analysis and public service, what users really need is not just a text, but a quick understanding of the relationship between data. Therefore, Baidu attempts to convert complex information into charts, cards and interactive interfaces using AI smart visualization and professional chart templates to make the data " visible, comparable and interactive ".
This also means that the search results began to change from " content pages " to " application pages ". Users do not have to jump to other websites or tools before computing, filtering and comparing, and more and more operations can be done directly on the outcome page.
Baidu states that this set of GUI capabilities will continue to be accessible to the ecology and to authoritative data institutions and professional partners.
Put AI interactive content into the stream
The content side is also an important direction for this upgrade of Baidu APP. Baidu formally launched the AI Interactive Theatre and AI Interactive Application, which seeks to move content from one-way consumption to user engagement.
AI Interactive dramas allow users to select different branches at key playpoints, thereby changing the course of subsequent stories. For example, users can decide on their own the key choices in a story, and different choices can enter a completely different scenario. AI interactive applications are closer to light quantization applications and small games that combine text, role, knowledge and play.
Baidu wishes to recommend direct access to the information stream for such interactive content. The search is responsible for understanding the issues raised by the user, while the recommendations judge the interest that the user has just generated in the browsing process and then push the relevant interactive plays and applications to the user. In the back, relying on the Baidu information stream for many days, Agent recommends that the instant interest of “I too want to try” be translated directly into next participation.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.