Google Antigravity 2.0 Enters the AI Coding Tool War

Google launched Antigravity 2.0 at I/O 2026 on May 19 — a substantial update to the company's agent-first development platform. New surfaces in this release: a desktop application, a command-line tool rewritten in Go, and an SDK that lets teams host Antigravity on their own infrastructure. The framing is multi-agent orchestration for serious coding work — not autocomplete inside an editor.

What shipped this week. The desktop application is the most visible change for individual users. Alongside it, Google has shipped a CLI rewritten in Go (faster start-up and a tighter terminal workflow) and a new SDK that supports running Antigravity inside your own infrastructure with optimization for Gemini models. The SDK is the part that probably matters most for teams: agent execution stays inside your network rather than routing through a Google-hosted endpoint, which changes the cost and security calculus for organizations that can't ship their codebase through a third-party service.

Multi-agent orchestration is the centrepiece. Antigravity 2.0's headline capability is dynamic subagents — a parent agent can spawn additional agents in parallel for branching tasks, then coordinate their results. The platform also supports scheduled and background tasks: code reviews, refactors, test runs, or repository hygiene work that fires on a schedule instead of waiting for user input. Security features added in this release include cross-platform terminal sandboxing, credential masking, and what Google calls "hardened Git policies" — guardrails to prevent agents from accidentally destructive operations on protected branches.

Google's headline demo. At the developer keynote, Google demonstrated an Antigravity-driven workflow that the company says built "the core framework of an operating system in roughly 12 hours" using 93 subagents, 2.6 billion tokens, and under $1,000 in compute. This is Google's claim, not an independently reproducible benchmark. The demo's prompt, success criteria, and methodology have not been published in a form that would let outside teams replicate the result. Treat the figure as a directional signal of what Google wants to show is possible, not a measured productivity gain for everyday teams.

Where Antigravity 2.0 sits in the market. For builders choosing what to use this week, the fair framing is this: Antigravity 2.0 is a credible new competitor in the agent-first coding workflow — alongside Cursor, Claude Code, and GitHub Copilot Workspace. The differentiators are real: the SDK plus self-hosted execution path, the Go CLI, and the tightly integrated Google Cloud, Firebase, and AI Studio ecosystem. The unknowns are equally real: independent benchmarks from teams outside Google, week-to-week reliability with Gemini 3.5 Flash as the underlying model, and how well multi-agent orchestration actually holds together on messy real-world codebases. Don't read the launch as a verdict on the category. It is an entry into the race, not a finishing position. For where the existing two leaders compare today, see Cursor vs. Claude Code; for the broader landscape, our AI coding tools roundup covers GitHub Copilot, Windsurf, Cline, and the rest.

$100 in Antigravity bonus credits — expires May 25. The most time-sensitive piece for individual buyers: Google is offering $100 in Antigravity bonus credits to new and existing Google AI Ultra subscribers, with the offer expiring May 25, 2026. The credits are for Ultra subscribers only and terms apply (credit usage rules, expiry restrictions, and other conditions per Google's announcement — read those before relying on the credit). If you were already evaluating Ultra at the new $100 entry tier — see our AI Tool Pricing Database for how that compares to other AI subscriptions — the bonus raises the effective value of the first month meaningfully. The honest advice, though, is to not subscribe just to claim the credit. Subscribe if Ultra fits your workload, and take the bonus while you are there.

What to do this week. If you ship code daily and you are already on Cursor or Claude Code, do not switch immediately. Run Antigravity 2.0 in parallel on a real task — a multi-file refactor, a feature touching three or more modules, or a bug investigation in code you don't fully remember — and judge whether the multi-agent orchestration actually earns its place in your workflow. The SDK is the long-arc story: self-hosted agent execution changes what's possible for teams that need on-prem or VPC-scoped tooling. If you're earlier in the stack-selection process and not yet committed to an agent-first IDE or terminal, Find My Tool can help narrow the choices before you commit.