Antigravity: reviews and analysis

TLDR: Antigravity is the Google editor where the agent does not live in a chat window: it drives the editor, the terminal and a built-in browser, and when it finishes it hands back screenshots and recordings of what it did rather than a step log. The individual tier is listed at no charge. In exchange you accept two uncomfortable things, quotas that have been cut without warning and an editor that has slipped into the background of a bigger platform.

What Antigravity is and how it works

Antigravity arrived the day Google unveiled its latest frontier model, and the idea behind it is easy to state. An agent that can only write in the editor is working with one hand. To check whether what it wrote actually runs, it needs to execute things and it needs to open the app, so Antigravity hands it all three tools at once: the editor, the terminal and a browser inside the program itself.

Its second decision follows from that, and it is the one most people cite. When the agent finishes, it does not leave a wall of text describing its steps. It leaves artifacts: a plan, a task list, the diff and, if the task touched the interface, screenshots plus a recording of the browser doing what you asked. You comment on those artifacts the way you comment on a shared document, and the agent picks the comment up without stopping what it is doing. For work delegated in long stretches, that way of reporting back earns its keep.

Part of the story needs telling to understand what you are installing. Antigravity is no longer an editor, it is a platform of four pieces: the editor, an agent command centre, a terminal tool and a developer kit. The editor still downloads and still works, but it is the piece Google has stopped pushing, and that shows in how often new versions land.

What it is like day to day

On a bounded, well-described task the result convinces. The agent decomposes, executes, tests in the browser and comes back with material you can review without reading a thousand lines. People coming from asking a side chat for things describe the shift as going from dictating to handing out work.

The rough part shows up under sustained use. The tool does not display how much context you have burned and does not let you compact it by hand, so long sessions degrade without warning. Users report projects duplicating themselves until the list fills up, and storage ballooning after a handful of sessions. The agent also makes expensive mistakes when it over-reads a cleanup instruction, so it pays not to grant permissions it does not need.

The loudest episode was not technical but about how users were treated. An automatic update replaced the editor with the agent command centre on installations that were already running, wiped configurations and add-ons, and left plenty of people without the tool they had chosen. You can get the editor back, it downloads separately, but the trust of anyone who lived through that has not fully returned.

Integration is well served. It accepts models from vendors other than Google, speaks the standard protocol for wiring in external tools and brings the browser inside. The limit is extensions, which come from the open registry rather than Microsoft’s, so check the ones you use daily before moving.

Pricing and plans

The individual tier is listed on the official site at no charge, and that includes the full editor. It is the cheapest way into the segment, no argument.

The small print is in what is not published. Above the free tier, capacity depends on Google subscriptions and on a pool of credits, and the vendor never says how many tokens a credit is worth. Without that conversion, working out what a month of work will cost is impossible, and that opacity is the underlying complaint repeated in forums. Add quotas revised downward more than once without notice and you have the reason its real-price score is no better despite starting at zero.

Translated into a decision: as a free tool for trying agents seriously, saying yes costs little. As a tool to build a team’s work on, today you cannot budget for it.

Who it is for (and who it is not for)

Antigravity is for anyone delegating long tasks who wants to check them in pictures, for anyone already paying for a Google subscription who wants to use it, and for anyone curious about agents without paying to start.

It is not for anyone who needs to know what they will spend. Kiro has a plan ladder with actual figures there, and Cursor an understandable subscription. It is also not for anyone depending on specific Microsoft ecosystem extensions, nor for anyone who wants an editor that feels instant, which is Zed territory.

Antigravity alternatives

If what draws you is delegating and reviewing, Kiro covers that same need with a published price. If you want editing speed, Zed. If you want the most polished AI integration, Cursor. They sit in Antigravity alternatives, ranked in our guide to the best AI coding tools.

Frequently asked questions

Is it really free?

The individual tier is listed at no charge on the official site and gives access to the full editor. What is neither free nor unlimited is agent capacity, governed by quotas and credits, and that is where the simple part of the answer ends.

How much does a credit spend?

Nobody knows, and that is the criticism. The vendor does not publish how many tokens a credit consumes, so there is no way to estimate the cost of a month’s work before starting it.

Does the editor still exist, or only version 2.0?

It still exists and downloads separately. What changed is its place in the house: the platform has four pieces and the editor is no longer the one Google pushes, with new versions arriving further apart than the rest.

Can I use models that are not Google’s?

Yes. It accepts third-party models and you can assign a different one to each agent inside the same session, which is among the most useful things it brings.

Do my VS Code extensions work?

Those in the open add-on registry do. Microsoft’s exclusive ones do not. That is the check to run before moving your whole workflow.