Google Antigravity 2.0: Features, Pricing, Benchmarks and How It Works (2026)

Google just unveiled Antigravity 2.0 at I/O 2026 on May 19 — and this is not an incremental update. It is a ground-up rebuild that replaces the original Antigravity IDE with something fundamentally different: a standalone, agent-first desktop application built for orchestrating multiple autonomous AI agents.

Here is everything we know — the model behind it, core features, pricing, benchmarks, and how it stacks up against the competition.

🔍 What Is Google Antigravity 2.0?

Antigravity 2.0 is not an IDE. Google has moved away from the traditional code-editor model entirely.

Instead, it is a standalone desktop application (macOS, Windows, Linux) designed as a command centre for AI agents. You define tasks, the agents execute — in parallel, on a schedule, or on-demand.

Think of it as a control room for autonomous coding agents — not a text editor with AI bolted on.

Old Antigravity vs. Antigravity 2.0

	Antigravity 1.0 (Nov 2025)	Antigravity 2.0 (May 2026)
Type	IDE (code editor)	Standalone desktop app
Approach	AI-assisted coding	Agent-first orchestration
Agents	Single agent	Multi-agent parallel
Scheduling	❌ None	✅ Cron-like tasks
Subagents	❌ None	✅ Dynamic spawning
Model	Gemini 2.5 Pro	Gemini 3.5 Flash
CLI	Gemini CLI	Antigravity CLI (Go)
SDK	❌ None	✅ Full SDK

⚡ The Engine: Gemini 3.5 Flash

Powering Antigravity 2.0 is Google's brand-new Gemini 3.5 Flash model, also announced at I/O 2026. Despite the "Flash" branding — which previously signalled a lightweight, budget model — this one is a genuine frontier performer.

Benchmark Results

Benchmark	Gemini 3.5 Flash	What It Measures
Terminal-Bench 2.1	76.2%	Real-world terminal/CLI tasks
MCP Atlas (Agentic)	83.6%	Multi-step agent workflows
SWE-Bench Pro	55.1%	Real GitHub issue resolution
CharXiv Reasoning	84.2%	Complex reasoning chains

Speed Advantage

This is where Gemini 3.5 Flash truly stands out:

280–455 output tokens/second — roughly 4× faster than comparable frontier models
Up to 12× faster within Antigravity thanks to platform-specific optimisations
For agentic workflows (tool calls → file reads → iterations), this speed compounds dramatically

Pricing Breakdown

	Price
Input tokens	$1.50 / 1M tokens
Output tokens	$9.00 / 1M tokens
Cached input	90% discount
Context window	1,048,576 tokens (1M)
Max output	65,536 tokens

🛠️ Core Features

1. Multi-Agent Orchestration

The headline feature. Run multiple agents in parallel, each working on different parts of your project simultaneously:

🔧 Agent A → building backend API endpoints
🎨 Agent B → creating the frontend UI
🧪 Agent C → writing and running tests

All three work concurrently, communicating results through the orchestration layer.

2. Dynamic Subagents

A primary agent can spawn specialised subagents on the fly. Need to resize an image, run a database migration, or lint a specific file? The main agent delegates to a subagent, keeps its own context window clean, and picks up the result when the subagent finishes.

This is significant — it solves the "context window pollution" problem that plagues single-agent tools during long coding sessions.

3. Scheduled Tasks (Cron-Like Automation)

Antigravity 2.0 introduces a scheduling layer that turns agents from interactive assistants into persistent automation pipelines:

🔒 Nightly security audit — scan for vulnerabilities every night
📚 Documentation refresh — update API docs after each merge
📦 Dependency checker — flag outdated packages every Monday
🧹 Code cleanup — automated linting and formatting on schedule

4. JSON Hooks

Intercept and control agent behaviour with JSON-based hook definitions — no complex wrapper code needed. Enforce project-specific rules like:

Banning certain dependencies
Enforcing naming conventions
Requiring test coverage for new functions

5. Voice Support

Integrated live voice transcription powered by Gemini Audio models. Talk to your agents instead of typing — useful for brainstorming, code reviews, or when you are away from the keyboard.

🌐 The Full Antigravity Ecosystem

Google shipped more than just a desktop app. The I/O 2026 announcement included an entire ecosystem:

Antigravity CLI

A terminal-based agent runner built in Go for speed and responsiveness. Shares the same agent harness as the desktop app, so your configuration and hooks work across both environments.

Antigravity SDK

Programmatic access to the agent harness. Build and host custom agents on your own infrastructure. Useful for:

Custom CI/CD agent pipelines
Internal developer tools
Specialised domain-specific agents

Managed Agents (Gemini API)

Spin up agents in isolated Linux environments with a single API call. No infrastructure to manage — Google handles the sandbox.

⚠️ Important: Google is retiring the legacy Gemini CLI for consumer users on June 18, 2026. If you use the Gemini CLI or its IDE extensions, start planning your migration to the Antigravity CLI now.

💰 Pricing and Availability

Antigravity 2.0 is available now (as of May 20, 2026) on macOS, Windows, and Linux.

Plan	Price	Antigravity Limits
Free tier	$0	Limited usage (exact quotas TBC)
AI Pro	Existing plan price	Standard usage limits
AI Ultra	$100/month	5× higher usage limits

🎁 Launch Offer

Google is offering $100 in bonus credits for new and existing AI Ultra subscribers until May 25, 2026 — essentially a free first month at full capacity.

Enterprise customers can access the platform through the Gemini Enterprise Agent Platform on Google Cloud.

⚔️ How It Compares to the Competition

Antigravity 2.0 enters a crowded field. Here is how it positions against the other major players:

Feature	Antigravity 2.0	Claude Code	OpenAI Codex
Architecture	Multi-agent orchestration	Single-agent, deep reasoning	Sandboxed managed agent
Speed	280–455 tok/s (fastest)	~60–80 tok/s	~100 tok/s
Context	1M tokens	200K tokens	128K tokens
Scheduling	✅ Built-in cron	❌ None	❌ None
Subagents	✅ Dynamic	❌ None	❌ None
Best for	Parallel workflows, automation	Complex refactoring, large codebases	CI/CD integration
Platform	Desktop + CLI + SDK	Terminal CLI	API-only

Bottom line: If your workflow benefits from parallel agents and recurring automated tasks, Antigravity 2.0 is the strongest option today. Claude Code remains the best for deep, single-pass reasoning on complex problems. OpenAI Codex fits best as a headless CI/CD agent.

🚀 Getting Started

Here is the quickest path to get up and running:

Download Antigravity 2.0 from antigravity.google
Sign in with your Google account
Start small — run a single agent on an isolated task to learn the interaction model
Try subagents — delegate file operations or test runs to subagents
Set up scheduling — automate one recurring maintenance task (linting, dep updates)
Scale up — run multi-agent workflows on larger projects

Tip: The learning curve is real. This is not "autocomplete on steroids" — it is a fundamentally different way of working with code. Give yourself a few days to adjust to the agent-first workflow.

Final Thoughts

Google Antigravity 2.0 is the most ambitious developer tool announcement at I/O 2026. The shift from "AI-assisted IDE" to "agent-orchestration platform" signals where Google believes development is heading — a future where you define intent, agents execute, and you review results.

The Gemini 3.5 Flash model's speed advantage is genuine and directly impactful for agentic workflows where the model needs to call tools, read files, and iterate many times per task.

If you are already using AI coding tools, Antigravity 2.0 is worth trying — especially with the $100 credit offer running until May 25. If you are new to AI-assisted development, simpler single-agent tools might be a gentler starting point before jumping into multi-agent orchestration.

Either way, the post-IDE era is officially here.

Last updated: May 20, 2026