Google I/O 2026 Unveils Gemini Agent Era: New AI Models, TPUs & Multimodal Tools

On May 20, Google’s I/O developer conference announced a comprehensive move into the "Gemini" agent era, unveiling a suite of AI hardware, models, and product upgrades.

Hardware and TPU Advances

Google introduced its eighth‑generation Tensor Processing Unit (TPU) with a dual‑chip architecture that separates training and inference. The training chip, TPU 8t, delivers roughly three times the raw compute of the previous generation and can mobilize over one million TPUs across global data centers. The inference chip, TPU 8i, focuses on latency reduction and generation speed, demonstrated by a live demo that generated a simple game at about 1,500 tokens per second.

Gemini 3.5 Flash and Omni Flash

The new Gemini 3.5 Flash model surpasses the previous 3.1 Pro in coding, agent capabilities, and tool usage. In the GDPval‑AA benchmark, it achieved 1,656 Elo, far above 3.1 Pro’s 1,314 Elo, while being four times faster than other leading models. Pricing is $1.50 per million input tokens and $9.00 per million output tokens, a 40 % reduction versus 3.1 Pro. The model’s knowledge cutoff is January 2025 and it supports a 1 million‑token context window.

Google also released the multimodal Gemini Omni Flash, the first model in the Omni family, capable of generating arbitrary content from any input modality.

Agents and Developer Platform

Agents are the core across all announced products. The Antigravity developer platform received a 2.0 update, replacing the old CLI with a globally available Antigravity CLI and allowing developers to deploy Google’s internal Agent Harness on their own servers. Combining Antigravity with Gemini 3.5 Flash yields a 12× speed boost.

In a 12‑hour demonstration, 93 sub‑agents processed over 15,000 model requests, handling 2.6 billion tokens and constructing a functional operating‑system core that can run commands, display animations, and even launch a Doom‑style game—capabilities not possible with Gemini 3.1 Pro.

Consumer‑Facing Agents

The consumer‑grade Gemini Spark (Google’s “OpenClaw”) acts as a personal assistant. At work, Spark can scan documents, emails, and chats to draft weekly reports in the user’s tone. In daily life, it can organize a neighborhood party, track RSVPs, enforce community rules, and generate polished presentation slides.

To support Spark’s cloud usage, Google introduced a $100‑per‑month “Ultra” plan offering five‑times the usage quota, 20 TB of storage, and priority access to Antigravity. The previous top‑tier plan was reduced from $250 to $200 per month.

Android and UI Evolution

Android now features “Halo,” an on‑screen agent monitor that shows real‑time agent activity, progress, and required user confirmations. The future UI is envisioned to serve agents rather than traditional apps.

The Gemini App was rebuilt with a new design language called “Neural Expressive.” Its Daily Brief feature reads your email and calendar each morning, extracts key items, and suggests next actions.

Search and Commerce Overhaul

AI‑Mode search, now serving over a billion monthly active users, supports multimodal inputs (images, files, video) and integrates conversational follow‑ups. Users can create “Search Agents” that monitor specific metrics (e.g., biotech stock indicators) 24/7 and push only relevant alerts.

Google introduced the Universal Commerce Protocol (UCP) and Agent Payments Protocol (AP2) to standardize AI‑driven shopping. UCP acts as an HTTP‑like layer for agents, while AP2 adds safety checks (brand, product, spend limits) with tamper‑proof digital authorizations. The Universal Cart leverages these protocols to enable cross‑merchant AI shopping, automatically finding discounts, checking inventory, and verifying component compatibility (e.g., CPU‑motherboard matching).

Creative Tools and Content Verification

New creative tools include Google Pics for precise image editing, Stitch for real‑time voice‑driven UI design, and Google Flow for AI‑assisted video and music generation. Flow’s “Music” feature can transform a piano recording into a polished R&B track.

Google’s SynthID watermark technology now tags over 1 trillion images, videos, and 60 k years of audio, allowing Chrome users to right‑click any image to verify authenticity. OpenAI and other companies have joined the SynthID initiative.

XR Hardware and Scientific Applications

Android XR smart glasses are launching in two lines: a lens‑display version later this year and an audio‑only version this autumn, designed by Gentle Monster and Warby Parker, manufactured by Samsung, and compatible with both iOS and Android.

Gemini for Science bundles a hypothesis generator, discovery engine, and scientific skill set, compressing complex pharmaceutical data analysis from hours to minutes. The Weather Next model accurately forecast Hurricane Melissa three days in advance, providing critical evacuation time.

In security, Code Mender automatically detects and patches vulnerabilities, while Google’s Isomorphic Labs accelerates AI‑driven drug discovery for immunology and oncology.

Overall, Google’s AI stack—from next‑gen TPUs to agent‑centric applications—demonstrates a pervasive shift toward AI‑driven workflows across everyday digital experiences.