Google Antigravity 2.0: agentic dev suite with desktop app, CLI and SDK

At Google I/O 2026 on 19 May 2026, Google introduced Antigravity 2.0 — an updated version of its agentic development platform. What used to be a single desktop app becomes a suite of three parts: a revamped desktop app, a new CLI tool and an SDK for custom workflows. Antigravity originally shipped on 18 November 2025 alongside the Gemini 3 model and competes with agentic coding tools such as Cursor. The real heart of the announcement is less any single feature than a strategic bet: Google opens its own IDE to rival models.

What used to be the case

Antigravity launched in November 2025 as a standalone desktop application — an agent-first IDE where the agent, not the editor, sits at the centre. Instead of typing line by line, you describe a task, and one or more agents plan it and carry it out across editor, terminal and browser. According to Google, users can spawn, orchestrate and observe multiple agents working asynchronously across different workspaces. Each agent can plan, write code, run terminal commands and use the web to verify its own results.

Even in that first version, model choice was the differentiator: from the start, Antigravity offered not only Google’s own Gemini but also models from Anthropic and OpenAI. What was missing was a way to use the same agent logic outside the graphical interface — in the terminal or embedded in your own scripts.

What’s the case now

With Antigravity 2.0, Google shifts the tool from an app to a platform with three entry points:

1. The desktop app orchestrates more agents in parallel. The revamped app puts running multiple agents simultaneously front and centre, rather than working through them one after another. Users can design custom subagent workflows and schedule tasks that run automatically in the background. Alongside this, Google has added native voice commands, much like those rolled out in Gmail and Docs.

2. A new CLI tool brings the agents to the terminal. Written in Go, the CLI is, according to Google, faster and more responsive than its predecessor and lets you launch agents without opening the editor. Important for existing users: the Antigravity CLI fully replaces the previous Gemini CLI. Anyone who built workflows on it has to migrate.

3. An SDK opens the platform to custom agents. With the Antigravity SDK you can build your own agents on top of Google’s coding tool and wire them into different infrastructure. This is complemented by a Managed Agents feature in the Gemini API that gives agents an isolated Linux environment to run in.

Model optionality remains: according to Google’s developer blog, Antigravity runs with generous rate limits on Gemini 3 Pro and fully supports Anthropic Claude Sonnet 4.5 and OpenAI GPT-OSS. On availability, Google opts for a low barrier to entry: a free public preview at no cost for individuals, usable with a private Google account, without a credit card and without a waitlist. Those who need more quota can subscribe to the new AI Ultra plan at $100 per month, which according to Google offers five times higher limits in Antigravity than the Pro plan. Many of the new features are powered by Gemini 3.5 Flash, also unveiled at I/O.

Context

Three of the loudest players in the market are pulling in the same direction: Google with Antigravity, the startup Cursor and Anthropic with Claude Code all bet that, going forward, multiple agents will work on a codebase in parallel. The question is no longer whether but how you steer a swarm of agents without losing control. Antigravity slots in here — what’s more interesting is how Google tries to set itself apart.

The bet is called model optionality. Unlike Anthropic, which in Claude Code naturally runs only its own models, Google deliberately allows Claude and OpenAI’s GPT-OSS inside its own IDE. That’s a notable move: Google sells the platform and the orchestration, not necessarily the model underneath. For agencies that use such tools day to day, this shifts the lock-in question. Commit to a platform that knows only one model, and you’re tied to that model’s pricing and quality trajectory. A model-agnostic interface promises the option to switch models when another becomes cheaper or better — without rebuilding the whole workflow.

The plan doesn’t fully add up: using Antigravity ties you to Google’s interface, its CLI and its SDK — and the new CLI replaces the previous Gemini CLI with no fallback, forcing existing setups into a migration. So the lock-in just moves up one level, from the model layer to the platform layer. And the generous rate limits explicitly apply to Gemini, not necessarily to the same degree for the rival models — so the economic incentive to ultimately run Google’s model after all remains. Model agnosticism here is a selling point, not a neutral referee.

What you can do now

If you’re evaluating tools for agentic coding: test Antigravity in the free public preview alongside Cursor and Claude Code — with no credit card, the barrier is low. Focus less on demo speed than on how well you can steer and observe several agents at once.

If model choice matters to you: check concretely how Claude Sonnet 4.5 and GPT-OSS behave inside Antigravity — and whether the rate limits there are as generous as for Gemini 3 Pro. This is exactly where it’s decided whether the model optionality is real or merely nominal.

If you built on the Gemini CLI: plan the migration to the new Antigravity CLI, since the predecessor is fully replaced. It’s not an emergency, but it won’t happen on its own either.

Entdecke mehr

Blog

Saving Tokens with Claude: 6 Principles That Make Experts Twice as Fast

How I turned my CLAUDE.md from a style guide into a token budget — 6 principles for lower cost, less waiting, and more honest reporting.

Glossar

Claude Code

Claude Code is Anthropic's official AI coding tool — available as a terminal CLI, VS Code extension, JetBrains plugin and web app, with a strong focus on tool use and configurable workflows.

Lexikon

AI Coding Tools Compared

Cursor, Windsurf, Claude Code, GitHub Copilot, Continue.dev, Aider compared. With table and decision guide for four typical developer workflows.

Themenuebersicht

KI-WerkzeugeAI-IDE