My Gemini 3 Review

Everyone else is going to obsess over benchmark numbers. They're going to do that because the numbers are, frankly, insane... truly wild improvements across the board. But I'm not going to do that here. I've been living with Gemini 3 for the past few days; actually working with it, building things, writing things, seeing what it feels like in practice. The benchmarks might tell you what it can do (and it can do a lot); I want to tell you how it feels to use.

The Model

Let's start with the creative writing, because that's where Gemini 3 first floored me. GPT-5.1, which dropped last week, was already a noticeable jump from previous frontier models. But Gemini 3? It wrote book chapters I had to double-check weren't plagiarized from a real book. The voice was coherent. The pacing natural, the turns of phrase genuinely surprising. But most importantly, it didn't feel like the "AI slop" writing we all know just a little too well. It's really impressive... Gemini 3 doesn't just put out "good for AI" writing, it puts out genuinely good writing.

The improvements feels fundamental. Previous models had a certain spikiness... their quality varied wildly depending on the task. You could get brilliance on one task, followed by just-okay results on another. Gemini 3 is more consistent, less prone to those jarring spikes. My hunch is that Google has cracked something about reinforcement learning on non-verifiable tasks... creative work where you can't just check if the answer is right. The result is a model that feels more like a skilled collaborator than anything we've had before.

That said, here's an important note: for 80% of your daily work, you might not even notice the difference. Current models are already "good enough" for writing emails or making small changes to your webapp. So at first glance, Gemini 3 doesn't always feel like a massive leap. But that feeling is deceptive. The jump is there, it’s just hiding in the difficult 20%... the complex reasoning, the subtle creative choices, the edge cases where other models fall apart. When you really need that extra brainpower, it's there.

Another standout feature: it's fast for how smart it is. To understand this, we can think of a metric like "intelligence per second", and Gemini 3 is fantastic in this regard.

I probably shouldn't compare it to GPT-5 Pro directly since their Deep Think (~equivalent mode) wasn't available for early testing, but, impressively, the regular version of Gemini 3 often outperformed GPT-5 Pro. Outperforms it, and does so without a 5-10 minute wait. You get both the quality and the speed, which changes how you work.

Personality-wise, it's a shift. Out of the box, Gemini 3 is less... ingratiating than most other models. It doesn't open with flowery praise, followed by three paragraphs of preamble. It's more terse. Direct. It gives you the answer and (mostly) stops. I prefer this. I don't need an AI to give me every little detail (if I need this, I'll ask); I need it to get to the point. With GPT-5.1, for example, I find myself scrolling through verbose explanations hunting for the actual content. Gemini 3 respects your time.

Other models have default "personas" and styles (UI, writing, etc.) that are very hard to escape; Gemini 3 just... listens and does what you ask. For example, if you prompt it to "write this like a cynical 1940s detective, but make it modern", it nails the specifics without fighting you and reverting back to the slop-styles we all know and hate.

The Antigravity IDE: Great, But Keep Your Eyes on It

The Antigravity IDE is impressive for a launch product. It feels like a real development environment, not a demo. The browser integration for testing sites is genuinely useful... it'll spin up a server, check if it acheived the goal it was working on, iterate without context-switching or human input. It's great.

But here's the thing: you have to babysit it. The model will sometimes glance at a log, declare victory, and move on while your build is still throwing errors. It'll screenshot a UI, say "looks good," and miss that the site wasn't even running in the first place. You need to keep the terminal open, re-run checks, and explicitly tell it to verify its work. Custom instructions help... "Keep reading the logs as you spin things up until you know it works." is a good one to add. For developers who stay engaged, it's powerful. For those wanting a magic button, it'll frustrate. That said, these are likely temporary issues that'll be patched via prompt updates on Google's side over time.

The Tradeoff

If GPT-5.1 is a solid junior engineer, Gemini 3 is a senior engineer who says "got it, done", and you better check that it's actually done. I keep reaching for it, not because it's perfect, but because when it's right, it's brilliantly, almost humanly right.

This is, without a doubt, my new daily driver. And with Google's computing power and ability to serve this cheaply and stably I'd bet this is going to be a winner.