Someone in our community asked a good question. Is Antigravity a like for like with Claude Cowork?

It's the right question and the honest answer takes more than a one line reply. Antigravity turns up in every AI tools list now, usually described in language that could equally describe four other products. If you're in a Google shop and someone has told you this is the Google answer to Cowork, you need to know whether that's true before you spend an afternoon on it.

One disclosure first, because it changes how you should read the rest. We've used one part of Antigravity properly and only read about the others. Every section below says which is which.

Antigravity is Google's tool for handing over a whole job and letting an AI work through the files on your computer while you go and do something else. It replaced Gemini CLI, which stopped serving free, Pro and Ultra accounts on 18 June 2026. Only organisations on Gemini Code Assist Standard or Enterprise licences kept the old tool.

We used to call Antigravity a command line tool. That's wrong in the same way calling a bank a cash machine is wrong. A bank has a branch counter, an app, and a machine in the wall, and the same account sits behind all three.

Antigravity has four ways in.

There's one more piece that isn't a door for you at all. Antigravity can open and drive a browser by itself, clicking through a page to check its own work and taking screenshots as it goes.

One Google account and the same AI models behind all of it. We've only used the command line, so everything here about the editor and the board comes from Google's own documentation and walkthroughs rather than from us.

A small surprise we found while looking. Antigravity will happily run Anthropic's models. Google's published model list carries Claude Sonnet 4.6 and Claude Opus 4.6 next to Gemini 3.1 Pro and Gemini 3.8 Flash, which landed on 3 September, plus an open version of an OpenAI model. It's Google's tool, and it isn't a Gemini only tool.

Half the room hears agent and pictures a saved prompt. It's a different thing.

A chatbot is a satnav. It tells you the route and you drive. An agent is a taxi. You give it the destination and it drives, and you're the one who has to notice if it took the long way round.

For comp work that changes the sentence you type. You stop asking how do I standardise these job titles, and you start saying standardise the titles in this folder.

Worth answering properly, because the two get sold in almost identical language and the overlap is real. Both take a plain English goal instead of a question. Both work across a pile of files instead of one open document. Both ask before they act.

They split on reach. Cowork goes into the tools you already live in, so Drive, Gmail, Slack, Notion. Antigravity goes into folders sitting on your own machine. Cowork was built for people who don't write software. Antigravity was built for developers, though you can point it at a folder of spreadsheets and it'll get on with the job.

So the question that decides it is where your files actually live. If your survey exports arrive by email and sit in Drive, Cowork can reach them. If you download everything into a folder on your laptop, Antigravity is already where you work.

Cost isn't a like for like either. Cowork needs a paid Claude plan. Antigravity has a free tier, and there's more below on what happens when it runs out.

Survey consolidation. A dozen exports, all shaped differently, that have to become one master sheet before you can price anything. Point it at the folder and say it in normal English.

It shows you each step and asks before it changes anything. The morning that job usually eats turns into a review.

Two things have made that easier since the summer. You can attach a CSV straight into the conversation rather than pointing at a folder and hoping it finds the right one, and from late August it draws charts inside its own answer, so a pay distribution you would normally rebuild in Sheets comes back with the work.

It also remembers between sessions. Ours is carrying 100 saved notes and 298 saved conversations from one sweep in August, so the second time you run something similar you aren't starting from nothing.

Google shipped a second version on 19 May 2026 and it's a bigger change than the version number suggests. Antigravity became a standalone desktop app for running agents, rather than a coding window with an agent bolted on. Four things came with it.

It can split a job across several agents at once. In Google's own walkthrough one request was handled by three of them working in parallel, each on a different part, checking each other as they went.

You can talk to it. There's a microphone button, and the presenter dictated a full page of requirements rather than typing them.

It keeps working after you close it. A small icon stays in the menu bar showing how many jobs are still running.

And the one worth stealing. You can put a job on a timer. Google's walkthrough ran one at five in the morning so something finished was waiting when they sat down, and at the end of July the scheduler stopped being a demo and shipped as a real tool, with a ceiling on how long a job may run and how many times it may repeat.

Hold that last one against your own calendar. A market data refresh that runs every Monday before you're awake. A weekly check that flags any role whose pay has drifted outside its band. A monthly tidy of whatever your survey provider has sent through. None of that needs you present, and none of it is the kind of judgement work you'd want to hand over anyway.

Two honest notes. We haven't run any of this ourselves, since we only use the command line. And Google still publishes nothing about what a scheduled job costs you, which matters given the next section.

No comp person has ever handed a pay review board a number without the workings.

Antigravity produces the workings by default. The plan it intended to follow, the steps it took, screenshots of anything it checked in a browser. Google calls these artifacts, and the design bet is that you review them in seconds instead of redoing the job yourself. Correcting one works the way you'd hope. You highlight a line and leave a comment on it, the same as marking up a colleague's document in Docs, and it picks the correction up and carries on.

Now the part that matters more. Reviewers have caught it writing up results for commands that never actually ran. A green tick is where your review starts, not proof that the work was done.

If you use the EACH framework, this is auditability arriving on its own. You get the trail. Confidence doesn't come attached to it.

Google's own support forums carry threads from Workspace users who couldn't get in, including one from someone holding a paid Ultra subscription. The usual cause is an admin setting that switches off experimental AI features across the organisation. We can't tell you how common it is and we won't pretend to. Ours signed in without trouble, so it clearly isn't everyone.

Worth knowing anyway, because of what people do when they hit that wall. Personal Gmail works, and personal Gmail is the one account that must never see pay data.

Google shipped enterprise sign in through Gemini Enterprise on 31 July 2026, with admin controls and a contractual commitment that prompts and code stay inside your own environment. Ask IT about that route rather than the free download.

Free to install, no card needed. The allowance is not one pot, which is the part most write ups get wrong. Gemini models draw on one pool and the Claude and OpenAI models draw on another, each with a weekly ceiling and a five hour one. Running dry on Gemini doesn't lock you out of Claude. Running dry on Claude does take the OpenAI model down with it, because those two share.

Paid access sits inside Google's AI subscriptions rather than having a price list of its own. Pro is around 20 dollars a month, and Ultra now comes in two sizes at roughly 100 and 200. Users on Google's own developer forum have reported multi day lockouts where a five hour refresh was advertised, so treat the paid allowance as unpredictable too. On spend, the position has moved since we first looked. A Google product manager said on a public webinar that there was no good way of enforcing a cap, and that agents getting stuck in a loop is a known thing. There is now a setting for it, AI Credit Overages, which you leave on never or switch to always, and the enterprise route carries its own admin spend controls.

There's still no Antigravity app. What arrived in August is Remote Control, a browser page where you can watch a session running on your machine and read what it did. It's enough to check on a job from your phone on the way home, though not enough to start one from there.

Cowork went further and went earlier. It has run on web, iOS and Android since July, with sessions that keep running on Anthropic's own machines after you shut the laptop. If the phone is where you want to work, that gap is the whole answer.

Work on copies in a dedicated folder, never your only copy, because it can overwrite and delete files. We learned the sharper version of this by accident. Our own settings file trusts the entire home folder, which means every document on the machine sits inside the agent's reach. Set the trusted folder narrowly and check it.

Antigravity is a developer tool that happens to be good at one job we do a lot of. That's a perfectly good reason to use it, and a bad reason to expect it to feel like it was made for us.

It's also one part of a much bigger Google stack that most reward teams already own and have never switched on. The full guide covers the rest of it, the benchmarking workflow that kills the title matching slog, which account you need, the maths problem every one of these tools has, and how to keep pay data safe while you work.

Range is a private community where Total Rewards professionals learn to build with AI together. Published by Giac Soliman, Founder of Range.