Google relaunched its flagship coding product at Google I/O 2026 on May 19, 2026, and the headline is the kind of thing you read twice: Google Antigravity 2.0 has no IDE.
Antigravity 1.0, which shipped in November 2025, was a fork of Visual Studio Code with an AI agent panel bolted on. Antigravity 2.0 throws the code editor away entirely. It is now a standalone, agent-first desktop application where you talk to agents, watch them work, and review what they produce. No file tree as the main event. No editor pane. The bet Google is making: the future of software work is *supervising agents*, not *editing files*.
This post covers what actually shipped, what it costs, how it compares to Claude Code and Cursor, and why the launch-day reception was bumpy.
What Antigravity Is, For People Who Missed 1.0
Google Antigravity is Google's "agent-first" development platform. Instead of writing code line by line with an AI helper, you give an AI agent a goal. It plans, writes, tests, browses the web, and edits files on its own while you supervise.
Version 1.0 launched in November 2025 as the Antigravity IDE — a modified VS Code fork released alongside Gemini 3. It had two surfaces in one app: a normal AI code editor, and an "Agent Manager" screen that hid the editor so you could focus on managing agents.
Antigravity 2.0 is a fundamental restructure. In Google's own words from the announcement:
"Google Antigravity 2.0 is a new, standalone desktop application that fully delivers on a truly agent-optimized experience, available on macOS, Linux, and Windows... Users interact with powerful agents both synchronously and asynchronously, and there is no IDE."
Google's stated reasoning: coding agents are expanding into general knowledge work, and bundling an IDE with an agent surface in one app is "confusing and potentially daunting to those less familiar with code and IDEs." They wanted to "rearchitect the product to be agent-first from the ground up, independent of an IDE or other dev-specific concepts such as repositories."
The old Antigravity IDE is not killed outright — it auto-updates to 2.0, and users are asked whether they want to keep the IDE around ("recommended for developers"). Google's actual recommendation is to "dual-wield" Antigravity 2.0 alongside whatever IDE you already use.
What Shipped: Four Surfaces, Not One
Antigravity 2.0 is not a single app. It is a platform with four entry points sharing one "agent harness."
1. The desktop app (the flagship)
- ●Dynamic subagents — "the main agent can dynamically choose to define and invoke subagents to complete focused subtasks, thereby not polluting the main agents' context window and allowing for parallelism of work."
- ●Asynchronous task management — tasks and commands run in the background without blocking the main agent.
- ●Scheduled Tasks — "you can define crons to trigger the invocation of Antigravity agents on a predefined schedule. No longer do you need to manually invoke every agent."
- ●JSON hooks — intercept and control agent behavior with a simple JSON config.
- ●Projects replace workspaces — agent conversations are no longer tied to a single repository. A "project" can span multiple folders with its own permissions.
- ●New slash commands:
/goal(run to completion without asking for input),/grill-me(ask clarifying questions before starting),/schedule(one-time or recurring timer),/browser(explicitly control when the agent uses the browser). - ●Live voice transcription in the input box.
2. The Antigravity CLI
A new terminal companion — "the most lightweight way to invoke, monitor, and interact with Antigravity agents." It shares the same agent harness and settings as the desktop app. Conversations started in the CLI can be pulled into the desktop app.
It replaces the Gemini CLI (more on that below).
3. The Antigravity SDK
A Python library — pip install google-antigravity — that is Apache 2.0 licensed with a public GitHub repo. Google claims "a functional agent in under 15 lines." It supports MCP servers (stdio, SSE, Streamable HTTP), custom Python tools, agent skills, declarative safety policies, and nine lifecycle hook points. TypeScript and Go support are on the roadmap, along with Gemma integration for self-hostable open models.
4. Managed Agents in the Gemini API
"With a single API call, you can spin up an agent that reasons, uses tools and executes code in an isolated, ephemeral Linux environment." Custom agents are defined with markdown files (AGENTS.md, SKILL.md) rather than orchestration code.
The Model: Gemini 3.5 Flash, Served 12x Faster
Antigravity 2.0's default model is Gemini 3.5 Flash, which also launched on May 19, 2026. Google's framing:
"Don't let the name 'Flash' mislead you when it comes to raw model power for complex, agentic tasks."
Google says Antigravity agents built "a fully operational OS, using only Gemini 3.5 Flash." On speed:
"Normally, Gemini 3.5 Flash is 4x faster than other models with frontier intelligence. For a limited time, Antigravity is serving Gemini 3.5 Flash 12x faster (not just 4x faster), thanks to further inference tricks."
The "12x faster for a limited time" is a launch promotion, not a permanent number. Worth knowing before you build a workflow around the speed.
The Gemini CLI Is Dead
This is the underreported story. Antigravity 2.0's launch comes bundled with the retirement of the Gemini CLI:
- ●May 19, 2026 — Antigravity CLI available to everyone.
- ●June 18, 2026 — Gemini CLI and Gemini Code Assist IDE extensions stop serving requests for free users, Google AI Pro, and Ultra subscribers.
- ●Enterprise exception: organizations on Standard/Enterprise licenses keep Gemini CLI access.
Get the Weekly IT + AI Roundup
What changed this week in NinjaOne, ServiceNow, CrowdStrike, and AI. One email, every Monday.
No spam, unsubscribe anytime. Privacy Policy
Your Agent Skills, Hooks, Subagents, and Extensions carry over to the Antigravity CLI as plugins, and Google ships migration guides. But the practical reality is: if you built workflows on the Gemini CLI, you have a 30-day window to migrate.
This is Google consolidating a famously fragmented dev-tool lineup — Gemini CLI, Gemini Code Assist, Project IDX, Firebase Studio, Jules — into one "Antigravity" brand with a shared agent harness. Google's stated goal: "a single, cohesive developer product offering that leverages a shared agent harness co-optimized with Gemini models."
Pricing and Availability
Antigravity 2.0 is generally available now — free download for macOS, Windows, and Linux. No waitlist for the desktop app or CLI. The SDK and Managed Agents are in preview. There is no web version.
| Plan | Price | Antigravity limits |
|---|---|---|
| Free tier | $0 | Generous rate limits |
| Google AI Pro | ~$20 / mo | Baseline |
| AI Ultra (new mid-tier) | $100 / mo | 5x higher limits than Pro |
| AI Ultra (top tier) | $200 / mo (cut from $250) | 20x higher limits than Pro |
Two pricing moves at I/O 2026: Google added a new $100 / mo Ultra tier and cut its top Ultra plan from $250 to $200. Note that Google's plan naming is genuinely confusing — press coverage refers to both a "$100 AI Ultra" and a "$200 AI Ultra." Verify the exact tier names and limits on Google's pricing page before committing budget.
How It Compares to Claude Code and Cursor
Three tools, three genuinely different philosophies.
| Antigravity 2.0 | Claude Code | Cursor | |
|---|---|---|---|
| Form factor | Standalone desktop app (no IDE) + CLI + SDK | CLI-first, terminal-native; also IDE extensions | Full IDE (VS Code fork) |
| Default model | Gemini 3.5 Flash (Google models only) | Claude Opus / Sonnet (Anthropic only) | Model-agnostic — you pick |
| Where agents run | Google-hosted Linux sandboxes (via Gemini API) | Locally, on your machine | Locally; cloud agents optional |
| MCP support | Yes, via SDK and CLI | Yes, native and first-class | Yes, one-click catalog |
| Pricing entry | Free tier; $20 / $100 / $200 | Bundled with Claude Pro ($20/mo) or Max; API pay-as-you-go | Free Hobby; $20 Pro; up to $200 Ultra |
| Autonomy | High — /goal runs to completion, Scheduled Tasks run unattended | High — configurable, auto-approve mode | Medium-high — agent mode with diff review |
| Target user | Google-ecosystem devs, and non-coders doing knowledge work | Terminal-native developers, automation builders | Developers who want a polished AI IDE with model choice |
The clearest contrast: Antigravity 2.0's flagship is a GUI desktop app built around watching agents work. Claude Code's flagship *is* the terminal. Cursor's flagship *is* the editor. Antigravity executes agents in Google-hosted Linux sandboxes; Claude Code executes locally by default. If you want to keep code on your machine, that distinction matters.
One more philosophical split: Antigravity is the only one of the three explicitly designed to expand past coding into general knowledge work. Google is not positioning this as just a developer tool.
The Rocky Launch
We always cover the launch reception honestly, and this one needs it.
Antigravity 1.0 had a poor reputation. The dominant complaint was rate limits — users reported 7-day lockouts after 20 to 30 minutes of use, and the community openly called the tool a "paperweight." There were also reports of silent throttling, opaque quota displays, agents crashing and freezing, and a prompt-injection vulnerability that Google patched in April 2026.
2.0's headline features — Scheduled Tasks, async agents, faster Flash, longer sessions — directly target 1.0's biggest pain points. But day-one reception suggests the stability problem is not solved. Launch-day Hacker News threads were dominated by crash reports rather than enthusiasm:
"Antigravity used to be rock solid for me, but now I can't even authenticate when it opens. 2.0 Bricked me."
A separate thread documented a real packaging bug: the 2.0 installer drops files into the existing IDE's directory, so both the IDE and the new app launch 2.0 — making the old IDE unreachable. That directly contradicts Google's tidy "two apps coexist peacefully" story.
Eagle-eyed readers also caught a typo in Google's own announcement: it says the Antigravity IDE launched in "November 2026," a date that has not happened yet. (It launched November 2025.) Minor, but not a great look on a flagship launch post.
None of this means the product is bad. It means the launch was rushed and the day-one build is buggy. If you depend on stability, wait two or three weeks for the first patch cycle before moving real work onto it.
Who Should Care
Try Antigravity 2.0 now if:
- ●You are deep in the Google ecosystem and want agents wired into Workspace, Firebase, and Android
- ●You want cron-scheduled autonomous agents and the desktop "agent cockpit" model appeals to you
- ●You are a non-coder doing knowledge work and an editor-free agent surface lowers the barrier
- ●You are on the Gemini CLI and need to plan your June 18 migration anyway
Wait, or stick with what you have, if:
- ●You need launch-day stability — the day-one build is buggy
- ●You want your code to stay on your machine — Antigravity runs agents in Google-hosted sandboxes
- ●You want model choice — Antigravity is Gemini-only
- ●You already have a working Claude Code or Cursor setup and nothing here is a must-have yet
The Bigger Picture
Strip away the launch bugs and the real story is a strategic one. Google looked at its own sprawling, confusing dev-tool lineup and collapsed it into a single brand with a single agent harness. It then made a genuine bet — that the editor itself is becoming optional, and that the primary interface for software work is a window where you watch agents and approve their output.
That bet might be right. Claude Code, Grok Build, and OpenAI's Codex are all pushing the same direction, away from the editor and toward delegation. Antigravity 2.0 is just the most literal expression of it: Google removed the code editor from its coding tool on purpose.
Whether the execution catches up to the vision is the open question. Right now the vision is ahead.