Quick Answer
Antigravity 2.0 is Google's rebuilt AI coding platform, announced at Google I/O 2026. It drops the IDE shell entirely and ships as a standalone desktop app (macOS, Linux, Windows) built around autonomous agent orchestration, powered by the new Gemini 3.5 Flash model.
It matters because it is Google's most direct attempt yet to compete with Anthropic's Claude Code and OpenAI's Codex. Instead of bolting a chat panel onto a code editor, Antigravity 2.0 treats agents as the primary interface and lets one main agent spin off subagents that run coding work in parallel.
From IDE Add-On to Agent-First Desktop App
The original Antigravity launched as a coding assistant living inside an IDE-style framework. The 2.0 release strips that out. The new desktop app is a centralized home for agent interaction. You give the system a high level goal, and a main agent plans the work, defines subagents on the fly, hands them distinct tasks, and runs them concurrently.
The practical win is the asynchronous execution model. Code compiles and tests run in the background without freezing the main workspace or filling the context window with build logs. For long running refactors and migrations, that is the difference between watching a single chat thread crawl forward and having a small team of agents work in parallel while you review results.
Gemini 3.5 Flash: The Speed Story
Antigravity 2.0 runs on Google's newly unveiled frontier model, Gemini 3.5 Flash. Independent throughput numbers from Artificial Analysis place it well ahead of competing tools on raw output speed.
| Model / Tool | Output tokens per second | Core focus |
|---|---|---|
| Antigravity 2.0 (Gemini 3.5 Flash) | 289 | Parallel agent orchestration |
| OpenAI GPT-5.5 | 71 | General multimodal logic |
| Claude Opus 4.7 | 67 | Complex advanced reasoning |
Speed alone is not the full picture, but in agentic workflows it compounds. When a planning agent dispatches ten subagents, a 4x throughput advantage means the whole job finishes in a fraction of the wall clock time. Google's internal benchmarks also claim Gemini 3.5 Flash beats its predecessor Gemini 3.1 Pro on standard coding and agentic tasks.
To prove the point at I/O, Antigravity head Varun Mohan said the team instructed the agents to build an operating system from scratch and then code a playable Doom clone to run on it, spending under $1,000 in API token costs end to end.
New Slash Commands and Terminal Controls
The 2.0 release introduces slash commands that give developers explicit control over how aggressive the agents are.
- /goal. Run a complex, multi-step task to completion without checking in for confirmation.
- /grill-me. Force the agent to stop and ask clarifying questions about architecture before touching any file.
- /browser. Explicit toggle for web browsing. Google conceded that autonomous agents still struggle to decide when they actually need live external data.
The /grill-me command is the most useful addition for production codebases, where an agent guessing wrong about an interface can create hours of cleanup work.
The Wider Antigravity Stack
Beyond the desktop app, Google is shipping a unified stack:
- Antigravity CLI. Replaces the legacy Gemini CLI. A lightweight, high velocity terminal product for developers who live in the shell.
- Antigravity SDK. Lets founders host the same agent harness on their own infrastructure using Markdown configuration files.
- Google AI Studio mobile app. Coming soon. Export entire prototypes from your phone straight into a local Antigravity production environment in one click.
Restructured Pricing
Google has redrawn its subscription tiers to bring in enterprise power users without forcing them onto the flagship plan.
| Plan | Monthly price | Usage vs Pro |
|---|---|---|
| Pro | $20 | Baseline |
| AI Ultra | $100 | 5x higher Antigravity limits |
| Flagship Ultra | $200 (down from $250) | 20x Pro limits |
New and existing Ultra subscribers can claim $100 in promotional bonus credits inside the app before May 25, 2026.
How It Compares to Claude Code and Codex
Claude Code still leads on dense reasoning tasks where a single, careful agent outperforms many fast ones. Codex remains tightly integrated with OpenAI's broader product line and is strong on multimodal prompts. Antigravity 2.0 is making a different bet: that most real engineering work is parallelizable, and that a fast model with strong orchestration beats a slower model with deeper single-thread reasoning.
That bet will not hold for every codebase. For systems where one wrong refactor cascades into broken builds, you probably still want Claude Opus driving. For greenfield builds, prototypes, and large parallel migrations, Antigravity 2.0 looks like the strongest option on the market today.
The Takeaway
Antigravity 2.0 is Google going on the offensive in developer tooling. The agent-first desktop app, the 289 tokens per second throughput on Gemini 3.5 Flash, the new SDK and CLI, and the redrawn pricing tiers all point at the same goal: own the workflow, not just the model. Whether it converts Claude Code and Codex users will depend less on benchmarks and more on whether teams trust autonomous agents to ship production work.
