Gemini 3 and Google Antigravity launch

Published on

The long-awaited Gemini 3 Pro from Google and Google DeepMind is FINALLY out!

Being a member of the Google AI Builders community, I got early access to the model, and my fingers were itching to talk about it.

First of all, the benchmark numbers are scary (in a good way). I’m leaving all of them below, but I’ll highlight the 37.5% on “Humanity’s Last Exam” without tools and the 31.1% on “ARC-AGI-2”, hitting double the competitors on what are maybe the most complicated benchmarks for LLMs today.

Gemini 3 benchmarks

Gemini 3 Deep Think benchmarks

It’s also worth highlighting the huge improvement on agentic programming benchmarks. In my initial tests, the results were truly excellent in zero-shot.

Besides being the “year of agents”, 2025 has also been the “year of agentic IDEs”, and along those lines, they also announced “Google Antigravity”, a new IDE for agentic development competing with VS Code, Cursor, Windsurf, etc. It’s already available for download, and from what they showed in the video, you can not only ask for code, but interact with specific parts of screens generated by Nano-banana (the image generator) to request changes (a mix of Canvas and Lovable), and collaborate with specific agents that can run in the editor, in the terminal and in your browser to plan and execute tasks.

I’ll test both Gemini 3 Pro and Google Antigravity more in the coming days and weeks to do a more detailed analysis (keep an eye on the IA Sob Controle podcast too). But either way, we’ve got another HOME RUN from Google with this model. It was basically looking at the competitors and saying: “Follow the leader.”