What AI Updates Did Google Announce at I/O 2025?

TL;DR

Google’s central announcement was a faster expansion of Gemini across Search, Meet, mobile devices, browsers, developer tools, and agent-based workflows. Gemini 2.5 Pro led cited model benchmarks, while Google also introduced Beam for realistic 3D communication, real-time translation in Meet, broader Gemini Live access, Project Mariner improvements, and an Agent Mode capable of researching options and scheduling appointments.

Transcript

[MOUSE CLICKING] [MUSIC PLAYING] - This town ain't seen nothing yet. [NEW RADICALS, "YOU GET WHAT YOU GIVE"] [APPLAUSE, CHEERING] SUNDAR PICHAI: Wow, cool video. Looks like the team has been having some fun. Hello, everyone. Good morning. Welcome to Google I/O. [APPLAUSE, CHEERING] So good to see everyone here in Shoreline. And hello to everyone jo... Read More

Key Insights

• Gemini 2.5 Pro is presented as Google’s leading general-purpose model, ranking first across all cited LMArena categories and reaching the top position on WebDev Arena. The updated model improved its WebDev Arena score by 140 Elo points over the previous version.
• Ironwood is Google’s seventh-generation TPU, designed to support thinking and inference at scale. Google says it delivers ten times the performance of the previous generation, provides 42.5 exaflops of compute per pod, and will reach Google Cloud customers later in the year.
• Google’s AI usage is growing rapidly across products and APIs. Monthly token processing increased from 9.7 trillion to 480 trillion within a year, while more than 7 million developers built with the Gemini API through Google AI Studio and Vertex AI.
• AI Overviews are reaching more than 1.5 billion users each month, making Search a major distribution channel for Google’s generative AI. Google identifies AI Mode as the next significant step for Search, alongside the continued expansion of AI Overviews.
• Google Beam is an AI-first video communication platform derived from Project Starline. It combines six camera perspectives into a realistic image on a 3D light-field display, with millimeter-level head tracking and real-time rendering at 60 frames per second.
• Real-time speech translation in Google Meet is designed to preserve a speaker’s tone, speech patterns, and expressions while translating conversations. English and Spanish support is available to subscribers, additional languages are planned within weeks, and enterprise availability is expected later in the year.
• Gemini Live now includes Project Astro’s camera and screen-sharing capabilities, allowing users to discuss objects, scenes, or content visible on their devices. Google says people are applying these abilities to activities including interview practice and marathon training.
• Project Mariner is a research prototype that lets AI agents interact with websites and perform user-controlled tasks. It can oversee as many as ten tasks simultaneously, while its Teach and Repeat feature learns a plan after a user demonstrates a task once.

Questions & Answers

Q: What were the main AI announcements at Google I/O 2025?

Google highlighted Gemini 2.5 Pro, the Ironwood TPU, Google Beam, real-time speech translation in Google Meet, Gemini Live camera and screen sharing, and improvements to Project Mariner. It also introduced Agent Mode for the Gemini app and announced Gemini SDK compatibility with MCP tools, supporting a broader ecosystem in which agents can access services and complete user-controlled actions.

Q: How much did Google’s Gemini usage grow before I/O 2025?

Google reported processing 480 trillion tokens per month across its products and APIs, compared with 9.7 trillion one year earlier, which it described as roughly fiftyfold growth. More than 7 million developers had built with the Gemini API, Gemini usage on Vertex AI had risen more than fortyfold, and the Gemini app had surpassed 400 million monthly active users.

Q: What is Google Beam and how does it work?

Google Beam is an AI-first video communications platform developed from Project Starline. An array of six cameras records a participant from different angles, and an AI video model combines those streams for presentation on a 3D light-field display. The system provides head tracking down to the millimeter and renders at 60 frames per second in real time.

Q: How does real-time translation work in Google Meet?

Real-time translation in Google Meet translates spoken conversation while matching characteristics such as the speaker’s tone, speech patterns, and expressions. Google introduced English and Spanish translation for subscribers, with more languages scheduled to arrive in the following weeks. The company also said the capability would become available to enterprise customers later in the year.

Q: What can Gemini Live do with Project Astro technology?

Gemini Live uses Project Astro’s camera and screen-sharing capabilities to discuss what a user can see on a device or through its camera. This lets users ask questions about visible surroundings and on-screen content during a live conversation. Google cited interview practice and marathon training as examples and began rolling out the features to everyone on Android and iOS.

Q: What improvements did Google make to Project Mariner?

Project Mariner gained multitasking support that allows it to oversee as many as ten tasks simultaneously. It also received Teach and Repeat, a feature through which a user demonstrates a task once and the system learns a plan for similar future tasks. Google planned to provide Mariner’s computer-use capabilities to developers through the Gemini API more broadly during the summer.

Q: How does Agent Mode in the Gemini app complete tasks?

Agent Mode combines Gemini’s reasoning with tools that can research information and take actions under user control. In Google’s apartment-search example, it finds listings that satisfy budgets and amenity requirements, uses Project Mariner to manipulate specific website filters, accesses listing services through MCP, and can schedule a property tour after the user identifies an apartment worth visiting.

Q: How is Google supporting communication between AI agents and services?

Google is supporting an agent ecosystem through multiple complementary technologies. Its open Agent2Agent protocol is intended to let agents communicate with one another and launched with support from more than 60 technology partners. Google also made the Gemini SDK compatible with Anthropic’s model context protocol tools, allowing Gemini-based agents to access other services while performing user-directed work.

Summary & Key Takeaways

• Google presented rapid gains in Gemini performance, adoption, and infrastructure. Gemini 2.5 Pro led the cited LMArena and WebDev Arena rankings, while the new Ironwood TPU promised substantially higher performance. Monthly token processing reached 480 trillion, and more than 7 million developers had built with the Gemini API.

• Google announced several communication and assistant features derived from research projects. Google Beam converts multiple camera feeds into a realistic 3D video experience, Google Meet gains speech translation that preserves vocal qualities, and Gemini Live receives camera and screen-sharing abilities originating from Project Astro on Android and iOS.

• Google expanded its vision for practical AI agents through Project Mariner, Agent Mode, and interoperable developer tools. Mariner can manage simultaneous browser tasks and learn repeatable workflows. Gemini SDK compatibility with MCP tools and support for Agent2Agent protocols are intended to help agents access services, communicate, and complete user-controlled actions.