速报 趋势信号

Google I/O 2026 速报:
两小时发布会,这几件事跟你有关

今天下午刚结束的 Google I/O 2026 主题演讲,将近两小时,全是 AI。信息量太大,帮你筛过一遍,留下值得知道的。

📌 2026.05.19 · 来源:Google 官方博客 / 9to5Google / Engadget / TechRadar / Tom's Guide
9 亿
Gemini 月活用户,一年翻了一倍
9.7 万亿
Gemini 每月处理的 token 数
13
Google 超 10 亿用户的产品数

Sundar Pichai 上台第一句话就在讲数字。但数字不是重点,重点是 Google 今天在用这些数字铺垫一件事:Gemini 不再是一个聊天助手,它正在变成一个 24 小时替你干活的 agent。

新模型

模型 Gemini 3.5 Flash
比 3.1 Pro 更强,速度是竞争对手前沿模型的 4 倍。编码、agent 任务、多模态三个方向全面提升。今天开始在 Gemini app、Search AI Mode、Antigravity 2.0 和 API 里全面上线。
📅 今天起全球可用
模型 Gemini 3.5 Pro
正在测试中,下个月上线。Flash 是 3.5 系列的第一个,Pro 是更强的版本。
📅 下月
模型 Gemini Omni
「create anything from any input」——接受图片、音频、视频、文本输入,输出视频。先从视频生成开始,后续扩展到更多输出格式。
📅 今天起可用

Agent 层:今天最重要的发布

Agent Gemini Spark
24/7 私人 AI agent。跑在 Google Cloud 的虚拟机上,笔记本合上它也在跑。基于 Gemini 3.5 Flash,集成 Gmail、Docs 等 Workspace 应用,夏天扩展到第三方工具(通过 MCP)。可以帮你监控信用卡账单里的隐藏订阅、追踪孩子学校的邮件通知、从 Gmail 里整理项目笔记生成文档。执行敏感操作前会先问你确认。
📅 下周面向 AI Ultra 用户(美国)开放 beta
Agent Daily Brief
每天早上给你一份个人简报。从 Gmail、Calendar、Tasks 里筛选当天要做的事,按优先级排列,附下一步建议。
📅 今天起面向 AI Plus / Pro / Ultra 用户上线(美国)
Agent Information Agents(Search)
在 Google Search 里 24/7 后台运行,持续跟踪你关心的事——博客、新闻、社交帖子、金融数据、体育赛事——有更新就通知你。
📅 今夏面向 AI Pro / Ultra(美国)
Agent Android Halo
手机上看 agent 在干什么的入口。屏幕顶部显示当前 agent 任务状态,有消息会提醒。
📅 今年晚些时候

Spark 的定位跟我们之前介绍的 Hermes Agent 非常像——都是 always-on、后台常驻、帮你盯着事情。区别是 Spark 跑在 Google Cloud 上(你不用自己维护服务器),但数据也在 Google 那边。Hermes 跑在你自己机器上,数据完全本地。两条路,各有取舍。

开发者工具

开发 Antigravity 2.0
Google 的 agent-first 编程平台。桌面端、CLI、SDK 今天全球可用。基于 Gemini 3.5 Flash,12 倍速度提升。多个子 agent 可以并行编码和测试。现场演示用 Antigravity 几十秒造了一个能跑 Doom 的简易操作系统。
📅 今天全球可用
三家 Agent 编程工具对比
Antigravity 2.0
Claude Code
Codex
模型
Gemini 3.5 Flash
Claude Opus/Sonnet
GPT / Codex
形态
桌面 + CLI + SDK
CLI
桌面 + CLI
并行 agent
✓ 多子 agent
Agent View(新)
单会话
手机遥控
Remote Control
ChatGPT App
发布状态
今天全球可用
正式发布
Preview

价格变了

价格 AI Ultra 分两档了
新增 $100/月档位:Gemini app 和 Antigravity 用量是 Pro($20/月)的 5 倍,含 YouTube Premium、20TB 云存储。原来的 $250 顶配档降到 $200/月,用量是 Pro 的 20 倍。同时 Gemini app 从「每日提示词次数限制」改成「compute 用量计费」——按你请求的复杂度和使用的功能来算,不再按次数。

几个值得注意的产品更新

产品 Google Search 近 30 年最大改版
AI Overviews 和 AI Mode 合并成统一体验。搜索框用 AI 生成建议(不只是自动补全),支持图片视频输入。还能在搜索里直接用 Antigravity 生成迷你应用(仪表板、追踪器)。全球今天上线。
产品 Docs Live / Gmail Live / Ask YouTube
对着 Google Docs 说话就能创建文档(语音 brain dump)。Gmail 里用对话式搜索找邮件。YouTube 里用对话式搜索找视频内容。
硬件 Android XR 音频眼镜
Samsung、Warby Parker、Gentle Monster、XREAL 等多家合作。通过语音和耳机交互,不在镜片上显示信息。支持导航、翻译、拍照。今秋发布,支持 iPhone。
产品 SynthID + C2PA 扩展
AI 生成内容水印验证扩展到 Search 和 Chrome。OpenAI 也宣布会使用 Google 的内容验证标准。这是两家竞争对手在安全问题上的罕见合作。

这场发布会真正在说什么

两个小时讲了几十个产品,但主线只有一条:Google 在把 Gemini 从「你问它答」变成「它替你做」。Spark 是 24/7 agent,Information Agents 是后台监控,Daily Brief 是每日代办,Android Halo 是手机上看 agent 状态——所有东西都指向同一个方向:AI 不再等你来找它,它主动帮你盯着事情。

Anthropic 在做同样的事(Claude Code + Remote Control),OpenAI 也在做同样的事(Codex + ChatGPT 手机遥控 + Deployment Company)。三家的路径不同,但终点一样。

Demis Hassabis 结尾说了一句:「AGI is now on the horizon.」不管你信不信这句话,今天发布的这些产品已经在改变你跟 AI 打交道的方式了。

RECAP TREND SIGNAL

Google I/O 2026 Recap:
The Two-Hour Keynote Had One Real Story

The Google I/O 2026 keynote was nearly two hours of AI announcements. Here is the filtered version: what matters, why it matters, and how it connects to the agent shift.

Source: Google official blog / 9to5Google / Engadget / TechRadar / Tom's Guide / 2026.05.19
900M
Gemini monthly active users, roughly doubled in one year
97T
tokens processed by Gemini every month
13
Google products with more than 1B users

Sundar Pichai opened with numbers, but the numbers were not the point. They were the setup for one message: Gemini is no longer just a chat assistant. Google wants it to become a 24/7 agent layer that works across products.

New Models

Model Gemini 3.5 Flash
Google positioned Flash as stronger than Gemini 3.1 Pro, with about 4x the speed of competing frontier models. Coding, agent tasks, and multimodal work all get upgraded. It is rolling out in the Gemini app, Search AI Mode, Antigravity 2.0, and the API.
Available globally from today
Model Gemini 3.5 Pro
The Pro model is still in testing and is scheduled for next month. Flash is the first member of the 3.5 family; Pro is the stronger version coming after it.
Expected next month
Model Gemini Omni
Google described Omni as "create anything from any input": image, audio, video, and text inputs, with video generation as the first output mode. The broader promise is a model that can move across input and output types more fluidly.
Starting with video generation

The Agent Layer Was the Real News

Agent Gemini Spark
Spark is a 24/7 personal AI agent running on Google Cloud virtual machines. It is based on Gemini 3.5 Flash, connects to Workspace apps like Gmail and Docs, and will later reach third-party tools through MCP. Example tasks include monitoring hidden credit card subscriptions, tracking school emails, and turning Gmail threads into project notes. Sensitive actions still ask for confirmation.
Beta next week for AI Ultra users in the U.S.
Agent Daily Brief
A personal morning brief pulled from Gmail, Calendar, and Tasks. It prioritizes what needs attention and adds suggested next steps instead of simply listing events.
Rolling out to AI Plus, Pro, and Ultra users in the U.S.
Agent Information Agents in Search
Background agents inside Google Search can keep watching topics you care about: blogs, news, social posts, financial data, sports, and other live sources. When something changes, they notify you.
Coming this summer for AI Pro and Ultra users in the U.S.
Agent Android Halo
A phone-level entry point for seeing what your agent is doing. It shows current agent task status at the top of the screen and alerts you when the agent needs attention.
Expected later this year

Spark is close in spirit to Hermes Agent: always-on, background-running, and built to watch things for you. The tradeoff is location. Spark runs on Google Cloud, so you do not maintain servers, but your data sits with Google. Hermes runs on your own machine, so the data stays local. Two routes, two sets of compromises.

Developer Tools

Dev Antigravity 2.0
Google's agent-first coding platform now ships as desktop app, CLI, and SDK. It runs on Gemini 3.5 Flash, claims a 2x speed improvement, and supports multiple sub-agents coding and testing in parallel. In the demo, Google used Antigravity to build a tiny operating system that could run Doom.
Available globally from today
Agent Coding Tools Side by Side
Antigravity 2.0
Claude Code
Codex
Model
Gemini 3.5 Flash
Claude Opus/Sonnet
GPT / Codex
Form
Desktop + CLI + SDK
CLI
Desktop + CLI
Parallel agents
Yes
Agent View
Single session
Mobile control
-
Remote Control
ChatGPT App
Status
Available now
Released
Preview

Pricing Changed Too

Pricing AI Ultra now has two tiers
Google added a $100/month tier with 5x the Gemini app and Antigravity usage of Pro, plus YouTube Premium and 10TB cloud storage. The previous top tier dropped to $200/month with 20x Pro usage. The Gemini app is also moving from daily prompt counts to compute-based usage, where cost depends on request complexity and enabled features.

Other Product Updates Worth Noting

Product The biggest Search redesign in 30 years
AI Overviews and AI Mode are merging into one experience. Search suggestions become AI-generated rather than simple autocomplete, image and video inputs are supported, and Search can use Antigravity to generate mini apps such as dashboards and trackers.
Product Docs Live / Gmail Live / Ask YouTube
Google is pushing conversation into its core products: speak to Docs to create a document, search Gmail conversationally, and ask YouTube for specific video content.
Hardware Android XR audio glasses
Google named partners including Samsung, Warby Parker, Gentle Monster, and XREAL. These glasses use voice and earbuds rather than displaying information on the lens. Navigation, translation, and photos are the early use cases, and iPhone support is included.
Product SynthID + C2PA expansion
Google is expanding AI-content watermark verification to Search and Chrome. OpenAI also said it would use Google's content verification standard, a rare cooperation point between competitors on AI safety infrastructure.

What This Keynote Was Really Saying

Google announced dozens of products, but the main line was simple: Gemini is moving from "you ask, it answers" to "it does work for you." Spark is the 24/7 agent, Information Agents are background monitors, Daily Brief is the daily task layer, and Android Halo is the phone-level status surface.

Anthropic is moving in the same direction with Claude Code and Remote Control. OpenAI is moving there with Codex, ChatGPT mobile control, and deployment-oriented work. The routes differ, but the destination is similar.

Demis Hassabis ended with: "AGI is now on the horizon." Whether or not you believe that sentence, these products are already changing how people interact with AI.