The models in antigravity are way behind the competition. Are people seriously running this in production environments?
Is not the model itself, is the way they are reasoning…
for simple tasks they work and they cost less than competition (example I found is using an MCP to create diagrams on a cloud software, starting from text or images as input).
For complex tasks / programming, I completely agree with you, the product is useless / conterproductive. When it does an error, it starts hanging / looping around that error as its life mission changes from the main goal to correcting the error. I found that is super aggressive on context compression even if it has 1M token of context, it never goes beyond 150K token.