I Stopped Using Anti-Gravity (After Praising It)
And Switched Fully
A few weeks ago, I wrote about why Anti-Gravity was better than Claude Code.
I have to take that back. Sorry.
At the time, Anti-Gravity was more useful to me. And it did pretty well. But that did not last long, unfortunately.
What is Anti-Gravity
In case you don’t know. Anti-Gravity is Google’s own coding app which is based on Visual Studio Code. A full IDE. Integrated Gemini models. Visual awareness. Everything in one place.
It was fast, cheap, and powerful. And nice to use. At first.
But things changed. Pretty fast, actually.
Today, I don’t use Anti-Gravity at all anymore. Uninstalled it. Never looked back. Man, AI keeps changing fast. And I with it.
Now, I mainly use Claude Code (again). Every day. And recently, I’ve also started using Codex 5.3 in the terminal. Not the Codex app. Although that is pretty nice too.
Both have been far more reliable. Far more predictable. And honestly, just better.
Let me explain why.
The Crashes Started First
The first real problem I had with Anti-Gravity wasn’t the code quality. It was stability. We’ll come to code in a bit.
AG started crashing.
Not once. Not rarely. Regularly.
I’d be in the middle of implementing something. A React Native screen. A Flutter component. A backend endpoint. Even just a website part. And suddenly:
Agent error.
No useful explanation. No clear stack trace. No actionable message.
Just… broken. Again and again. Quitting the app didn’t do anything. Reinstalling didn’t matter. Updates didn’t change it. Even the new model update from a day ago to Gemini 3.1 Pro did not address this issue.
This is one of the worst things an AI coding tool can do. Because the entire point is flow. You’re working fast. Iterating. Shipping. And suddenly everything stops.
You restart the app. Sometimes it works again. Sometimes it doesn’t.
This becomes exhausting.
Claude Code almost never does this. It’s boring in the best way possible. It just works.
Boring is good. I like boring.
The Code Quality Got Worse
Then the code. This part surprised me a bit more. The benchmarks on reviews on Gemini Pro were pretty good.
Early on, Anti-Gravity with Gemini felt capable. Fast and smart enough for most tasks.
But over time, I noticed more problems. It generated code that looked correct — but wasn’t. Subtle bugs. More and more. No tests. No consistency. Even with the humongous 1 million token context window. Just worse than Claude with 250K token context.
And more:
Incorrect assumptions about APIs.
Weird architectural decisions.
Things that required more fixing afterward.
This defeats the purpose of using AI in the first place. The goal is to generate correct code quicker.
Claude Code is significantly more reliable here.
Especially with larger projects. Especially when refactoring. Especially when dealing with complex logic. And especially since Opus 4.6 came out.
It understands intent way better. Less guessing. More precision.
The RAM Usage Became Ridiculous
This was probably the most frustrating technical issue.
Anti-Gravity started using huge amounts of RAM on my MacBook. Not slightly more. Dramatically more.
Fans spinning.
System slowing down.
Other apps lagging.
For an IDE built on VS Code, not great. Especially one that’s supposed to improve productivity.
Claude Code, being terminal-based, is much lighter.
Now, I did also use Gemini CLI instead of AG. It did improve the RAM issues, of course. But the code and reliability didn’t improve.
The UI Actually Became a Weakness
This is ironic.
Originally, the UI was one of Anti-Gravity’s biggest advantages. A full IDE experience. Visual context. Integrated workflow. Just easier to use for non-techy people than a command line tool.
But over time, it just got clunky. Heavy. Slow. Poorly designed in subtle ways.
Gemini CLI Was Better
Again. I used Gemini CLI too. I didn’t immediately switch fully to Claude.
It did better than Anti-Gravity in some areas. More stable. Less heavy. Cleaner.
But the core problems remained. Code quality wasn’t as consistent. It still made strange mistakes. Still occasional errors. Still less predictable. And it doesn’t seem to get as much love from Google as AG does.
Gemini CLI is usable. But it is not the best tool.
Claude Code feels like the best tool.
That difference matters.
Claude Code Became My Default
Today, Claude Code is my main environment (again).
I use it for:
Web apps
Mobile apps (React Native and Flutter)
Backend systems
Refactoring
Planning architecture
Even writing tasks sometimes
It handles all of this extremely well. It understands larger systems better. It maintains context better. It makes fewer dumb mistakes.
This reduces cognitive load massively. Which is the real productivity gain.
Not speed. Clarity.
The Only Real Downside: Price
Claude is expensive. There’s no way around that. Compared to Anti-Gravity. Compared to Gemini. Claude costs more. Significantly more if you use it heavily.
This is its biggest weakness.
If price is your primary concern, Gemini tools are attractive. They’re cheaper. Sometimes the free plan is enough to do a lot of work.
And for many people, they might be good enough. I think Google is not even betting on Gemini being the best AI tool, but more so on the packages it comes with: All Google tools. And image and video tools are a lot better than competitors, to give them that.
If your workflow depends on code and on reliability, correctness, and stability, Claude justifies the cost.
It saves time.
And time is more valuable than subscription fees.
Then I Tried Codex 5.3
Recently, I started using Codex 5.3 in the terminal. Codex 5.3 is good.
Very good, actually. I am not the biggest fan of OpenAI and ChatGPT. But Codex is good.
Much closer to Claude Code than Gemini. It’s fast. Accurate. Stable.
And works extremely well in terminal workflows.
For my use cases — web apps, mobile apps, writing, general development — it performs almost at the same level as Claude Code.
Even faster, sometimes.
Terminal Tools Are Winning
I feel like AI is making the terminal cool again. Many people who have never touched CLI now do.
That’s cool.
I used to do a lot of terminal work years ago. But then nothing for a long time. It’s nice to be back.
The Bottom Line
Anti-Gravity was exciting. But it became unreliable. Crashes. Bad code. Excessive RAM usage. Poor UI decisions. Lackluster updates.
Claude Code is the easy winner for coding and organizational workflows.
It’s stable. Accurate. Efficient. And integrates perfectly into a serious development workflow.
Codex 5.3 is a strong alternative. Very close in quality. Very usable.
The only real downside of Claude remains the price and token limits.
But for now, the experience is worth it.