With the launch of Gemini 3, Google also announced Antigravity, their new VS Code clone focusing on Agentic AI workflows. Since it happens to be the best place to test out Gemini 3 (it is free during the initial rollout) and I’ve only used the terminal-based AI agents, I decided to give it a spin.
The setup was seamless - installation was easy and with Open VSX you get all the popular VS Code extensions. I decided I’d have it work on an incomplete personal project of mine. I connected it to my wsl folder and it opened up the project with the AI chat panel on the right. From a workflow perspective, it’s pretty similar to Claude Code but I quite liked Antigravity’s interface for planning - It creates task and implementation artifacts that you can annotate with feedback comments. The project itself was something I had tried vibe-coding with gemini-pro-2.5 but it struggled with the schema design and the interface was not mobile friendly. Gemini 3 one-shot this. I had the whole thing up and running with Render and Neon in under an hour.
I was quite impressed with the model but the experience of using Antigravity itself left much to be desired. For one, the supposedly generous free tier is quite stingy - I ran out of Gemini 3 (high) credits within the first 15 minutes of use and got about 40 minutes of Gemini 3 (low) usage. The other available options are Claude Sonnet 4.5 and GPT-OSS 120B - of which only Sonnet 4.5 is worth using. In addition, the UX is quite confusing - it’s designed to have multiple agents working in the background but you end up with situations wherein the model says something to the effect of excellent the feature is deployed - you can restart the server while the actual changes are still waiting for approval.
There’s also a separate Agent Manager interface that lets you run multiple background agents but it’s not as easy to use as Claude subagents. You have to start a conversation in a separate window and you can only use one model across all the agents. You cannot, for example, use Gemini 3 (high) for a complex background task while using a cheaper model for code completion.
Final verdict - I want to use Claude Code but with Gemini 3.