Google used its I/O 2026 keynote on May 19 to relaunch Antigravity as a standalone, agent-first development platform — and to retire the Gemini CLI in the process. What shipped is not an IDE assistant but a desktop application built entirely around agent orchestration, paired with a new CLI, an SDK, Managed Agents in the Gemini API, and an enterprise tier. Every surface defaults to Gemini 3.5 Flash.
What actually ships
Antigravity 2.0 exposes the same agent harness across multiple surfaces. The desktop app runs dynamic subagents for parallelized tasks, supports scheduled background automation that turns single-turn agents into persistent pipelines, adds native voice commands, and wires into Google AI Studio, Android, and Firebase.
The Antigravity CLI fully replaces the Gemini CLI. It shares the desktop's harness for feature parity and preserves Agent Skills, Hooks, Subagents, and Extensions — now called Antigravity plugins — while adding credential masking, hardened Git policies, and terminal sandboxing.
The Antigravity SDK gives programmatic access to the agent infrastructure for self-hosted deployments on customer-controlled hardware. Managed Agents in the Gemini API go the other way: a single call spins up an isolated Linux sandbox where an agent reasons, executes code, and calls tools with persistent state across sessions, available through the Interactions API and AI Studio. An Enterprise Agent Platform connects Antigravity directly to Google Cloud projects.
The model underneath
Gemini 3.5 Flash is the default across all of it. Google claims it beats the previous Gemini 3.1 Pro on almost every coding benchmark while running roughly four times faster than other frontier coding models — a speed gap that compounds across concurrent agent calls, which is the entire point of subagent orchestration. In a keynote demo, Google DeepMind's Varun Mohan showed Antigravity and Gemini 3.5 Flash building a functioning operating system from scratch in 12 hours.
Pricing and access
Google attached a $100/month AI Ultra plan offering 5x higher usage limits in Antigravity than the existing Google AI Pro tier. Agent definitions are authored as markdown files, with new templates in the AI Studio Playground. Alongside the platform, Google proposed WebMCP, an open standard letting browser agents call structured tools through JavaScript and HTML forms rather than screenshot-driven interaction, entering an origin trial in Chrome 149.
What changes for builders
This is Google consolidating fragmented surfaces — IDE plugin, Gemini CLI, API — into one harness aimed squarely at Cursor, Claude Code, and OpenAI Codex. The differentiators are infrastructure-level isolation, persistent sandboxed state, and scheduled background execution, not raw autocomplete. For teams already on Google Cloud, Managed Agents remove the sandbox-provisioning work that usually gates agentic deployments. The bet: developers who orchestrate fleets of cheap, fast subagents will outrun those running one expensive frontier model per turn.
