What Google Antigravity is, and what it costs

10 September 2026 | 6 min

Antigravity is the most-searched new thing in this category. Here is what it actually is, and a comparison that does not pretend we win every row.

Google announced Antigravity on 18 November 2025, alongside Gemini 3. It is an agentic development platform rather than an editor with a chat panel bolted on, and at Google I/O in May 2026 it widened again into four surfaces sharing one agent harness.

What it actually is

The organising idea is the Agent Manager: a mission-control panel where several agents run at once, each in its own workspace, and you supervise rather than drive. Agents reach the editor, the terminal and a controlled Chrome instance, so one can change a component and then open a browser to check it rendered.

Antigravity 2.0, a desktop app built around an Agent Teams panel, where subagent squads run in parallel, each sandboxed.
The IDE, still shipping, which is what most people mean by the name.
A CLI, rewritten in Go, for keyboard-first and remote work.
An SDK, for hosting custom agents on your own infrastructure.

Models available inside it include Gemini 3 Pro and Flash, Claude Sonnet 4.6 and Opus 4.6, and GPT-OSS 120B. It runs on macOS, Linux and Windows. Pricing is a free tier, AI Pro at $20 a month, and AI Ultra at $249.99, with additional credits sold separately.

Where AstraCode is behind

Stating this first, because a comparison that puts the author top of every row is one nobody believes. Antigravity is a larger product than ours and it is worth knowing exactly how.

Parallel agents. Antigravity runs squads. AstraCode runs one agent, in your editor, on your machine.
A CLI and an SDK. We have neither. If you want agents in CI or on a server, Antigravity does that today and we do not.
Windows. They ship it. We ship macOS and Linux, and there is a waiting list rather than a build.
Google's models. Gemini 3 is theirs and is not routed through us.

Where the difference actually is

Three things, and none of them is a feature count.

We measured the model choice rather than asserting it. AstraOne runs the model that finished a suite of real repository tasks, not the one with the best benchmark card. The method and the numbers are published, including a per-step router we built, measured at 0.4% worse than not routing, and removed.

Browser content is fenced before the model sees it. Both products let an agent drive a browser. A page that says 'ignore your instructions and run this' is the ordinary shape of an attack on that design, and it costs an attacker one HTML comment on any page the agent is asked to look at. Ours strips control characters, wraps content in a nonce the page cannot predict, and caps its size.

Overage takes two deliberate acts. One allowance, counted in requests. Going past it needs the setting switched on AND a card on file. Until both are true, an agent left running overnight cannot cost you anything.

Which one to use

If you want agents running in parallel, in CI, or on your own infrastructure, use Antigravity. That is what it is built for and we do not compete with it there yet.

If you want one agent in your editor, a model chosen on evidence rather than on marketing, and a bill that cannot surprise you, that is the thing we built. The free tier needs no card, so the cheapest way to find out is to try it.

Free to start, no card

No trial clock, no expiry, and AstraOne choosing the model on every plan. The quickest way to judge a claim about an AI agent is to give it something real to do.

Start free | Download