Owen

Posted on May 24 • Originally published at ofox.ai

Google Antigravity 2.0: Gemini's Agent-First Desktop Platform Explained

#ai #gemini #agents #tutorial

TL;DR: Google launched Antigravity 2.0 at I/O 2026 on May 19. It bundles a desktop IDE, the agy CLI, an SDK, a Managed Agents API tier, and an enterprise deployment path — five distribution shapes for the same agent runtime. Default model is Gemini 3.5 Flash. The old Gemini CLI stops serving free and Pro requests on June 18, 2026. Two parts actually matter: a persistent isolated-Linux sandbox inside the Managed Agents API, and the shift away from "IDE with AI bolted on" toward multi-agent orchestration as the default mental model. AI Ultra is the new $100/month tier; the top tier dropped from $250 to $200.

Antigravity 2.0 is not Google shipping a better IDE — it is Google admitting the IDE is no longer the unit of work.

What Antigravity 2.0 actually is

Antigravity 2.0 is Google's standalone agent development platform, distributed as five surfaces that share one agent harness. The 1.0 release in November 2025 was an agent-first IDE built on a VS Code fork. 2.0 keeps that IDE but stops treating it as the center of gravity, recasting the whole stack as a runtime for autonomous coding agents.

The five surfaces, in the order most developers will touch them:

Desktop app — A standalone manager for spawning and observing multiple agents in parallel, with native voice commands, scheduled background tasks, and integrations into AI Studio, Android, and Firebase.
Antigravity CLI (agy) — A terminal-first agent runner installed via curl | bash, sharing the same harness as the desktop app.
Antigravity SDK — Programmatic access to the harness so you can define custom agents and host them on your own infrastructure.
Managed Agents API — A tier inside the Gemini API where a single call spins up an isolated Linux environment with reasoning, tool use, and code execution. State persists across calls.
Enterprise — Distribution through the Gemini Enterprise Agent Platform with deployment guardrails for regulated buyers.

None of these are independent products with separate semantics. They are five distribution shapes for one runtime, and that is the architectural bet Google is making here.

What changed from Antigravity 1.0

Antigravity 1.0 (November 2025) was framed as an agent-optimized IDE. You opened a workspace, got an Editor View and a Manager Surface, watched agents work. Default model was Gemini 3 Pro, with optional Claude Sonnet 4.5 and GPT-OSS, and it shipped free on macOS, Windows, and Linux.

The substantive additions in 2.0:

Capability	1.0	2.0
Surface count	IDE only	Desktop + CLI + SDK + Managed API + Enterprise
Multi-agent orchestration	Manual via Manager Surface	Dynamic subagents for parallel workflows
Background execution	No	Scheduled tasks, native voice commands
Default model	Gemini 3 Pro	Gemini 3.5 Flash (co-developed with the platform)
Isolated execution	Local sandbox	Persistent isolated Linux per session, server-side
Pricing	Free preview	Bundled into AI Pro / Ultra subscription tiers

Gemini 3.5 Flash is the model bet. Google claims it outperforms Gemini 3.1 Pro across almost every benchmark while running roughly four times faster than other frontier models. Whether that holds outside Google's own evals is a separate question. The platform is engineered around it being true.

Installing the agy CLI in 60 seconds

The CLI is the fastest way to try Antigravity 2.0 without committing to the desktop app. Install:

curl -fsSL https://antigravity.google/cli/install.sh | bash

The installer detects your environment, drops a binary named agy (not antigravity) into ~/.local/bin/ on Unix or %LOCALAPPDATA%\Antigravity\ on Windows, and prints the PATH export line if needed. Add it to your shell profile if your shell does not pick it up automatically:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc
agy --version

Restart the terminal, run agy --version, and you are ready to spawn an agent. A typical first invocation looks like this:

agy agent run "refactor the rate-limit middleware to use a sliding window" \
  --repo ./services/api \
  --model gemini-3.5-flash

The CLI ships preview-quality on Linux. macOS and Windows are smoother right now — if you are on a Linux server and hit weird PTY behavior, that is expected. Agent Skills, Hooks, Subagents, and Extensions migrated over from the Gemini CLI as Antigravity plugins, so existing customizations carry forward.

The Managed Agents API is the actually new thing

The desktop app and CLI are polish. The Managed Agents API tier inside the Gemini API is the structural change.

A managed agent call gives you, in one HTTP request:

A reasoning loop powered by Gemini 3.5 Flash (or whichever model you select).
Tool use against built-in or user-supplied tools.
An isolated Linux environment that the agent can execute code in.
Persistent state — each interaction creates an environment that can be resumed in follow-up calls with files and process state intact.

That last point is the headline. Most agent APIs today are stateless: you call, the agent runs, the sandbox is destroyed. Managed Agents lets you treat the sandbox as a long-lived workspace the agent returns to across calls. For workloads like data exploration, long-running migrations, or any agent that needs to do real work over many minutes or hours without you babysitting a session, this removes a real category of plumbing.

If you have already wired up your own sandbox (Modal, E2B, Daytona, a homegrown Firecracker setup), Managed Agents is direct competition. If you have been putting it off because the operational lift is real, this is Google handing you the rented equivalent.

Pricing: what each tier actually buys you

Antigravity does not have its own pricing page. Usage is metered against the Google AI consumer plans:

Plan	Price	Antigravity allocation
AI Pro	Existing $20/month tier	Baseline usage included
AI Ultra	$100/month (new)	5× Pro limits
AI Ultra top tier	$200/month (down from $250)	20× Pro limits
Free (legacy)	$0 via Gemini CLI	Ends June 18, 2026
Enterprise	Gemini Enterprise Agent Platform	Custom contract

Two things to flag. First, the price drop on the top tier from $250 to $200 is the same week OpenAI quietly tightened the limits on its Pro tier — competition on the high-end agent subscription market is doing real work right now. Second, the June 18 free-tier cutoff is the load-bearing date for anyone using the old Gemini CLI in production scripts: requests will start returning errors, not silently downgrading.

How it stacks up against Claude Code, Codex, and Cursor

A short, honest comparison from the perspective of "I want an AI agent to do work in my repo":

Claude Code is the standard for code-quality-first agent work. It is single-agent by design, model-locked to Anthropic, and has the most mature hooks/subagents/skills ecosystem. See our Claude Code extensibility guide for what that looks like in practice.
Codex CLI is the OpenAI answer — model-locked to OpenAI, lightweight, BYO-everything. Our Codex CLI real-world workflow walks through a day with it.
Cursor is the agentic IDE for people who still want a strong editor surface. The interesting question with Antigravity 2.0 is whether Google's desktop app eats this lane.
Antigravity / agy is the first major release where multi-agent orchestration is the default UX rather than an advanced feature. If your workload is naturally parallelizable (multiple small tasks, multiple repos, agents that can work independently), this is the only platform built around that shape.

For a deeper four-way breakdown, the AI coding agents compared 2026 post covers Claude Code, Codex, Cursor, and DeepSeek TUI side by side — Antigravity is the obvious next entry in that bracket.

When Antigravity 2.0 is the right call

Pick Antigravity if any of these apply:

You are already inside Google's developer ecosystem (Firebase, AI Studio, Android) and want one auth surface.
Your workloads parallelize naturally — multiple small tasks across multiple repos, or a single task that decomposes cleanly into subagents.
You need persistent server-side sandboxes and do not want to operate them yourself.
You are an enterprise buyer who needs a single vendor for both the runtime and the model.

Skip it if:

You want model portability above all. The platform supports Claude and GPT-OSS, but you pay a real latency and cost penalty for not staying on Gemini.
You are happy with a single-agent loop and find multi-agent orchestration adds more noise than throughput. (Honest take: for most one-developer workflows, this is true.)
You want to keep your costs purely pay-per-token rather than a $100/month subscription floor.

The third bullet is where API gateways come in. If you want Gemini 3.1 Pro (currently on ofox), Claude Opus 4.6, and GPT-5.4 Pro behind one OpenAI-compatible endpoint without subscribing to anyone's flagship plan, our LLM API aggregation guide covers how to wire that up. For where Gemini 3.1 Pro itself sits on price and performance, the Gemini 3.1 Pro API pricing & performance guide has the breakdown.

For the broader landscape of which model to pick for which job, the Claude 4 vs GPT-5 vs Gemini 3 comparison guide and Best AI Model for Agents 2026 are the right starting points.

What this means for the broader agent platform race

Every major lab now has a vertically integrated agent stack. Anthropic ships Claude Code, OpenAI ships Codex CLI, Google ships Antigravity. The independent IDEs (Cursor, Cline, Aider) compete against all three. The model-agnostic middle is getting squeezed.

The free-to-paid transition is the other piece. Google moved its CLI from "open and free" to "bundled with a subscription," and the June 18 cutoff for the legacy Gemini CLI free tier is part of a clear pattern: free coding agents are loss leaders that get retired once a platform has the user base it wanted.

Look at the Managed Agents tier for where the API surface is going. Stateless completion calls are now the bottom of the stack. The next layer up is "give me a sandbox plus a reasoning loop and resume it across calls." Whoever owns that layer owns the developer agent market the way OpenAI's chat completions API owned the previous one.

If you are paying $100 a month for AI Ultra just to keep your agent loop running, the agent loop is not the bottleneck — your gateway is.

Sources

Originally published on ofox.ai/blog.

DEV Community