Sorry Google, Antigravity isn't ready to take on Claude Code and Codex

I've been really getting into vibe coding lately, that I've even been using local LLMs with Claude Code for my workflow. I wanted to try something new, so I tried what Google had to offer in the same place.

I have had a few weeks with Antigravity 2.0 now, and the demo and the day-to-day experience are two very different things. If you are someone who uses Claude Code or Codex as part of a real workflow, Antigravity is not ready to replace either of them.

Google's models just aren't good enough

Benchmarks aren't the real deal

I know the technically correct thing to say here is that Claude Code, Codex, and Antigravity are all just harnesses. They read your files, manage context, run commands, and hand the actual thinking off to a language model. The tool itself is not intelligence. That is true, and it matters.

But I think that framing lets Google off too easy. Because when you sit down and use one of these tools, you are not using a harness in isolation. You are using the whole package: the agent layer plus whatever model it defaults to. And in practice, most people are not swapping out models or pointing Antigravity at a third-party API. They are just using it as it ships. So that is how I tested it.

I gave all three the same task: build a Notion clone. Codex ran on GPT-5.5, Claude Code ran on Sonnet 4.6, and Antigravity ran on the new Gemini 3.5 Flash.

Let's start with Antigravity's attempt:

It does not look bad. You can tell what it is supposed to be, which is something. But here is the problem: nothing in the app actually works. I can type in the title header, and that is it. Every button, every interaction, completely broken. Clicking anywhere does nothing.

I gave it the benefit of the doubt and prompted it again to fix the issues. It did not really do anything. The app remained broken.

Then I looked at what Claude Code (left) and Codex (right) put up:

Night and day. Both apps worked on the first try. Codex had a slight edge on design and feature completeness, but Claude Code was not far behind. Neither of them shipped me a UI that was essentially a static screenshot.

I tried a few other things as well. I have a couple of apps published on the App Store, so I decided to implement a real feature across all three tools using Swift. Same story. Antigravity went nowhere and actually managed to break several other things that were already working. The other two handled it fine.

Google has tried to make up for the model gap with other features, and we will get to those. But when the output of your default model is a broken app on a task this common, the features are not really the conversation.

Antigravity has a unique feature... which is quite useless

It's cool, I guess

In typical Google fashion, the naming here is already a mess before you even open the product. There are three ways to use Antigravity: the CLI, the Antigravity app (which is the equivalent of the Codex app or the Code tab in Claude), and the Antigravity IDE. Three separate things, three separate entry points, and the branding does almost nothing to help you understand which one you actually need.

The IDE is the differentiator Google is pushing. It is a VSCode fork with integrations across Google's AI suite, and it gives you a more fleshed-out UI for the CLI with proper sub-agent controls built right in. On paper, that sounds interesting.

The problem is that Claude Code has a VSCode extension. Codex has a VSCode extension. And even if neither of them existed, nothing is stopping you from running Claude Code in a terminal pane inside your existing editor. You have always been able to do that. The IDE is solving a problem that was never really a problem to begin with.

It gets worse. Chats between the Antigravity app and the Antigravity IDE do not sync. Two products from the same company, built around the same agent, and your context does not carry between them. So if you start something in the app and want to continue it in the IDE, you are starting over.

Which brings me to the real question: what is Antigravity actually selling? The IDE is a VSCode fork with a unique feature that Claude and Codex mostly replicate through extensions. The app is just the CLI with a nicer face. And none of it talks to each other properly.

It beats both Codex and Claude Code at one thing

Google's better for a simpler life

There is one area where Antigravity pulls ahead, and I want to be fair about it because it is real. Token usage. For the same tasks, I was consistently burning roughly 30% fewer tokens with Antigravity than with Claude Code or Codex. It was also noticeably faster at finishing tasks. That is not nothing, and if cost is your primary concern, that gap is going to get your attention.

The reason is pretty straightforward. Gemini 3.5 Flash is a faster model built specifically for speed and efficiency. Google has leaned into that hard, and on that metric alone, it delivers. But at the same time, Google has been getting flak for nerfed token limits.

The problem is that token efficiency only matters if the output is actually usable. And as we covered, for anything beyond a simple task, it frequently is not. Google loves to lead with benchmarks, and Gemini 3.5 Flash looks great on paper. It just does not translate in real-world use the way those numbers suggest it should.

The moment you push it into anything intensive, a longer coding session, a multi-file refactor, anything that requires the model to actually hold a lot together, the quality tradeoff becomes the whole conversation. You are not saving tokens if you have to redo the work anyway.

It is kinda useful (sometimes)

Where Antigravity with Gemini 3.5 Flash actually makes sense is simple, low-stakes terminal tasks. Using it as an agent to do things on your computer, move files, run scripts, automate something repetitive. It is the fastest and cheapest way to do that short of running a local LLM. For that specific use case, it is genuinely the right tool.