Google I/O Unveils Gemini 3.5 Flash: Faster, Cheaper, Beats 3.1 Pro and Introduces Antigravity 2.0

At 1 am Google held I/O 2024, unveiling a suite of new AI products. The headline is Gemini 3.5 Flash, the new base model for all announced services.

Gemini 3.5 Flash is positioned as a “Flash” model that is fast and cheap. Benchmarks from Arena AI show 76.2 % on Terminal‑Bench 2.1, 1656 on GDPval‑AA agent tasks, and 83.6 % on MCP Atlas tool‑calling. Pichai said its output speed is four times that of other frontier models at 289 tokens / s. Pricing is $1.50 / M input and $9 / M output, compared with $2/$12 for the previous 3.1 Pro and $3/$15 for Claude Sonnet 4.6, making it the best‑value option.

The conference also introduced Antigravity 2.0, described as Google’s version of Claude Code. It is an agent‑centric development environment with three core capabilities: Dynamic Subagents (parallel child agents), Scheduled Tasks, and the AgentKit 2.0 SDK for custom agent behavior. Users can access it via a desktop app, a CLI, or an SDK. A demo showed an operating‑system kernel built in 12 hours, spawning 93 sub‑agents, processing billions of tokens for under $1,000.

Gemini Spark is a new mode in the Gemini app that runs persistent AI agents in the cloud, allowing tasks to continue when the user’s computer is off and to invoke Google services (Gmail, Docs, etc.) across applications. Future integration with third‑party tools via MCP is planned, and Android will get an “Android Halo” UI for monitoring Spark agents.

Gemini Omni, presented by DeepMind CEO Demis Hassabis, is labeled a “World Model”. It accepts any combination of text, image, audio, or video as input and can generate or edit arbitrary outputs, effectively editing the world rather than just creating it. Hassabis emphasized that the ultimate goal is to train robots in simulation before deployment.

The Search experience received the biggest visual upgrade in 25 years. AI Mode now defaults to Gemini 3.5 Flash and supports longer, multimodal queries. Two new capabilities are introduced: information agents that continuously monitor web sources, and generative UI that creates interactive visual tools or tables on‑the‑fly. The changes are framed as a defensive move to keep Search as Google’s primary entry point.

Overall, the I/O announcements position models as infrastructure and agents as the product layer, with cost‑effective, high‑speed Flash models, a unified agent development platform, and AI‑enhanced Search forming a coherent strategy.