Antigravity, just came out.
The rate limits for Gemini3 are unworkable. Your Ultra plan will get you a total of about three hours of coding time with it before it requires you to switch to Claude Sonnet 4.5 (I don't understand why non-google models are there), which is accessed on a low priority connection. So, good luck getting through.
Beyond that, they'll give you an unguarded baseline copy of a Gpt-oss instance, which seems like an odd choice. Why? Again, no idea.
What little we do get to see of Gemini3 is pretty good. Bedside manner needs work, but it's very competent, especially with layered back end systems and UI development. Has no trouble with Docker, which is nice.
One interesting thing about Gemini3: It has no patience for exposition, and has trouble walking through a complicated plan. It's okay, but not great for brainstorming. It doesn't contribute ideas or help you think through design problems, even when you ask it to.
If you talk for too long, it throws its arms up in the air and starts laying down the foundations of your project.
I find this entertaining.
The other thing I checked for is debugging. This version of Gemini can debug without having a nervous breakdown. So big improvements there over 2.5 pro.
It's great at following directions, and probably the better tool if all you want is competent output from a cognitively light or totally brain dead workflow. Probably also the tool to use if you already have a detailed plan written out, and you just want it implemented without the chitchat.
This is probably the best of the agentic vss forks. Smoother than Cursor with less hand holding.
I think I still prefer terminal clients like Claude Code and Opencode though.