Google I/O Introduces Gemini 3.5 Flash – Faster, Cheaper Than 3.1 Pro – and Antigravity 2.0

At 1 a.m. local time Google held its I/O developer conference, announcing a suite of new AI products. The headline is Gemini 3.5 Flash, positioned as a lightweight, high‑throughput version that replaces the previous flagship Gemini 3.1 Pro in most Google services.

Benchmark performance – In the Arena AI leaderboard Gemini 3.5 Flash achieved 76.2 % on the Terminal‑Bench 2.1 code benchmark, 1 656 on the GDPval‑AA agent task, and 83.6 % on the MCP Atlas tool‑calling benchmark. Detailed results are shown in the accompanying chart.

Speed and cost – Sundar Pichai (quoted in the keynote) said the model outputs tokens at four times the rate of other frontier models, reaching 289 tokens / s. Pricing is $1.50 per M input tokens and $9.00 per M output tokens, compared with $2.00 / $12.00 for Gemini 3.1 Pro and $3.00 / $15.00 for Claude Sonnet, making Flash the best price‑performance choice.

Antigravity 2.0 (Google’s Claude Code) – The new version adds three core capabilities: Dynamic Subagents (parallel child agents), Scheduled Tasks (background automation), and the AgentKit 2.0 SDK for custom agent behavior. Users can access it via a standalone desktop app, a CLI, or an SDK that can be embedded in their own servers. In the keynote Google demonstrated building a runnable OS kernel in 12 hours using Antigravity 2.0, spawning 93 sub‑agents, processing billions of tokens for under $1 000 total cost.

Gemini Spark – Spark is a new mode in the Gemini App that provides 24/7 persistent agents. Because the agents run on Google Cloud rather than the local machine, they continue work when the user’s computer is off and can invoke the full Google suite (Gmail, Docs, Sheets, Calendar, Chrome, Android). Future integration with third‑party tools via MCP (e.g., Canva, OpenTable, Instacart) and an Android “Halo” UI area for real‑time status are also announced.

Gemini Omni – Presented by DeepMind CEO Demis Hassabis, Omni is described as a “World Model” that supports any input (text, images, audio, video) to any output. Unlike the earlier Genie series that generated worlds for agents, Omni edits worlds, allowing multimodal transformations such as swapping characters or scenes in a video. Hassabis emphasized that the ultimate goal is to train robots in simulation before real‑world deployment.

Search upgrades – Google claims this is the biggest Search revamp in 25 years. Two new capabilities are added: (1) information agents that continuously monitor web, news, social, finance, shopping, and sports feeds; (2) generative UI that, when a complex query is issued, invokes Gemini 3.5 Flash and Antigravity to generate interactive visual tools, tables, or other UI elements on the fly. The generative UI will be free to all Search users this summer.

Pricing and availability – Gemini App now offers an Ultra tier at $100 / month (down from $250), while the top‑tier drops from $250 to $200. Gemini 3.5 Flash is the default model for the new AI‑Mode Search, and all announced products (Gemini App, Search, API, AI Studio) are immediately usable.

Strategic takeaway – The conference underscores Google’s view that large language models are now infrastructure, while AI agents are the primary product layer driving user engagement and cloud revenue.