Google I/O 2025: The Future Unveiled - Key Announcements

The Pro version steps up with "Deep Think," a sophisticated reasoning mode. This feature, built on cognitive modeling and parallel reasoning, allows Gemini 2.5 Pro to evaluate multiple hypotheses before delivering a response. This is particularly impactful for tackling complex mathematical problems and intricate coding challenges.

• Elo scores have surged by over 300 points compared to the first-generation Gemini Pro.
• Dominates the LMArena leaderboard across all categories, showcasing its superior performance.

Gemini 2.5 Flash focuses on speed and efficiency without compromising significantly on power. It boasts improvements across reasoning, multimodality, coding prowess, and handling long context information.

• Requires 20-30% fewer tokens for generating responses, making it more efficient.
• Enhanced capabilities in complex reasoning and multimodal understanding.

Priced at $249.99 per month and initially available in the US, the AI Ultra Plan is positioned as one of the most premium flat-rate AI services. This follows a trend seen with offerings like OpenAI's Pro subscription and Anthropic's Claude Max tier.

Key Features of the AI Ultra Plan:

• Delivers the "highest level of access" to Google's cutting-edge AI services.
• Includes access to the Veo 3 video generator.
• Provides access to the Flow video editing application.
• Offers a Terabyte (TB) of cloud storage across Google services.

AI Mode introduces enhanced shopping features, including agentic checkout capabilities, simplifying the online purchasing process directly from search results.

A new virtual try-on tool for clothing allows users to visualize how apparel might look on them, enhancing the e-commerce experience within Search.

Search Live in AI Mode and deeper integration with Google Lens provide more dynamic and multimodal ways to explore information, combining reasoning with real-world visual understanding.

• Ability to dive deeper with follow-up questions.
• Curated web links for further exploration.

Android XR represents Google's platform for extended reality, encompassing augmented reality (AR), virtual reality (VR), and mixed reality (MR). The showcased smart glasses aim to integrate digital content with the real world, potentially powered by advanced AI for contextual assistance.

• Samsung's Project Moohan headset, previously seen at Mobile World Congress, was highlighted, suggesting potential collaborations in the XR space.
• Focus on creating intuitive interfaces and practical applications for everyday use.

Beam technology is designed to create photorealistic 3D renders of callers in real-time, without the need for specialized glasses or headsets for the viewer. The goal is to make virtual conversations feel as if the person is physically present in the room.

• Creates an "uncanny sense of presence."
• Aims to be accessible without specialized viewing hardware.

Imagen 4 was unveiled as Google's most advanced image generator to date, boasting significant improvements in rendering text within images and overall visual fidelity.

The latest iteration of Google's AI video generator, Veo 3, promises enhanced quality, coherence, and control in generating video content from text prompts.

Google introduced Flow, a dedicated AI filmmaking tool. It ingeniously combines the capabilities of Imagen, Veo, and Gemini models to enable users to create cinematic scenes from simple text prompts, simplifying complex video creation workflows.

Launched in December as an early research prototype for human-agent interaction, Project Mariner has evolved with new capabilities:

• Multitasking: Ability to handle multiple tasks or goals concurrently.
• "Teach and Repeat" Method: AI learns from user actions and can replicate them.
• Supports open standards like the Agent2Agent Protocol.
• Compatibility with Anthropic's Model Context Protocol (MCP) via Gemini API and SDK.

Agentic (action-taking) AI features are becoming more prevalent across Google products, including Chrome, Search, and the Gemini app. Glimpses of Project Astra's capabilities, likely powering some of these advanced assistant features like Gemini Live, were showcased.

The models now support extended language capabilities in over 24 languages, featuring improved text-to-speech and more expressive, natural-sounding voices.

Key Audio Advancements:

• Support for audio-visual input, allowing models to understand combined audio and visual cues.
• Native audio out dialogue capabilities, available in a preview version via the Live API.
• Developers can build and fine-tune the tone, accent, and speaking style of conversational AI experiences.

Ironwood is the first TPU specifically designed to power "thinking" and inferential AI workloads at massive scale. This focus on inference is crucial for deploying responsive and intelligent AI services widely.

• Delivers 10 times the performance over the previous TPU generation.
• Achieves an astounding 42.5 exaflops of compute power per pod.
• This infrastructure strength enables Google to deliver dramatically faster models, even as model prices are significantly decreasing.

The vision for Gemini is an AI that is intelligent, deeply understands the context the user is in, and can proactively plan and take action across any device. Glimpses of this future were seen through demonstrations of Gemini Live, powered by aspects of Project Astra.

• Google is working towards building an AI with a deep understanding of physics and real-world interactions, which could lead to more intuitive and capable assistance.
• Emphasis on seamless operation and assistance across diverse devices and platforms.

Gemini 2.5 Pro is now recognized as the leading AI model on benchmarks like WebDev Arena and LMArena. Developers gain access to:

• Enhanced coding and web app building tools.
• An expansive up to 1 million token window for long context understanding, crucial for complex applications.
• Improved security against indirect prompt injections.

Developers will be able to leverage the agentic capabilities of Project Mariner through the Gemini API. Early access has been granted to partners like Automation Anywhere and UiPath, with a broader rollout anticipated in the summer.