Google announced Antigravity 2.0 at I/O 2026 on May 19, extending what was a single-surface product (the desktop agentic IDE that shipped November 20, 2025) into a five-surface platform. The desktop IDE remains, redesigned with multi-agent orchestration and native voice commands. Around it Google added an Antigravity CLI for terminal-first agent work (with a deliberate migration path from the older Gemini CLI), an Antigravity SDK for building custom agents on the same substrate as the IDE, a Managed Agents runtime tier inside the Gemini API where agents you author can run hosted, and an enterprise deployment path through Gemini Enterprise Agent Platform. AI Studio, Android, and Firebase integrations are wired in so projects move between them with an export tool. Pricing now sits at AI Ultra $100/month (5x the Pro limits) and Premium $200/month (20x Pro, down from $250). The IDE itself remains free for individuals in public preview, as it was in 1.0.
The architectural shape is the part of this release that matters more than the feature list. Antigravity 1.0 was an IDE: a place to author and run agents locally with model optionality (Gemini 3 Pro, Claude Sonnet 4.5, GPT-OSS). Antigravity 2.0 is the same authoring surface plus three production-ward surfaces (CLI, SDK, Managed Agents tier) plus one enterprise surface (Gemini Enterprise Agent Platform). The migration story is now coherent end to end: prototype an agent in the IDE, drop into the CLI for terminal-driven iteration, package it through the SDK, run it hosted in the Managed Agents tier when you need scale, and graduate to the enterprise platform when you need procurement and compliance. The cost of that coherence is substrate lock-in: the agents you write against the Antigravity SDK and run in the Managed Agents tier are not portable to other runtimes without rewriting. Google's bet is that the integrated path is worth the lock-in, especially with the Gemini 3.5 Flash co-development loop (the model that powers the platform was itself developed using the platform, which is the most honest dogfood signal a vendor can offer).
The ecosystem thread this fits into is the consolidation of "agentic IDE" as a category and the verticalization of the layer above it. A year ago there was Cursor, Windsurf, Claude Code, GitHub Copilot Workspace, and a handful of others, each treating "AI in the editor" as the product. The new shape (visible across Antigravity 2.0, the various Anthropic Claude Code expansions, and the OpenAI Codex moves into research and spreadsheets the same week) is that the IDE is the entry point but the contested layer is the runtime where the agent actually executes work, schedules background tasks, persists state, and hits production. Multi-agent orchestration in the desktop app, background-scheduled tasks, custom subagent workflows, voice commands, these are not just UX features, they are the surface that defines who owns the agent's operating substrate in your stack. If you accept the orchestration model in your IDE, the path of least resistance pulls you into the same vendor's CLI, SDK, and Managed Agents tier. The vendors who win this layer will not be the ones with the best editor experience, they will be the ones whose authoring surface seamlessly extends into the production runtime.
Monday morning, if you are already running Antigravity 1.0 locally, the upgrade question is whether to also adopt the CLI and SDK now or wait for production needs to force the choice. The free public-preview tier still covers the IDE, so the cost of trying the new CLI is low, but the SDK lock-in becomes real once you start writing custom agents against it. If you are not on Antigravity, the question is comparative: how does the five-surface platform stack against Cursor + your own runtime, against Claude Code + the Anthropic-hosted future, against rolling your own substrate. The Gemini 3.5 Flash dogfood story is a real signal of internal investment but does not by itself answer whether the substrate fits your workload. If you are evaluating multi-agent orchestration patterns, the Antigravity desktop app is now one of the cleanest references for what "custom subagent workflows with background scheduling" looks like at production polish, worth a tour even if you don't adopt it. And if you are tracking the agent-runtime consolidation thread, this release tightens it further: the IDE is now an entrypoint, the contested layer is everything downstream of it, and Google has now planted a flag across all five surfaces of that layer.