In my opinion this is the best way to utilize the antigravity models to make most of your tokens.
The Gemini 3.0 flash model evidently is only good at following instructions, it does have some capability to make decisions but they are not reliable.
Flash models are like Dev interns who don’t know what they fixed couple of commits ago, but you get a lot of them.
You need a good manager to handle them.
The Gemini 3.1 Pro is a good all rounder. It knows what it’s doing, the best model so far if you prompt it right.
The Pro can almost go unsupervised with minor bug fixes after review.
Pro is extremely limited even though you are paying for it.
My go to strategy to manage my output tokens and get the best possible result for Gemini models :
-
Use Flash models only for detailed feature implementation. If the feature is complex the flash model struggles to figure out the precise logic for implementing. It gives you random blocks of code connected together which is somewhat correct and if you try to fix it , it might make it better by breaking other things or it can make it worse, there’s literally no way to tell what it does.
The key is to prompt the model exactly what you want, something like increasing the resolution of the prompt. Best way is to prompt Gemini 3.1 pro chat what you want vaguely and it generates a detailed instruction that can be injected into the flash model in antigravity. ( This has worked out for me so far.) -
Use the Pro models only for feature improvement and bug fixes. it will do a great job of fixing bugs, improving the logic, optimizing etc. which will not take up most of your tokens as it is not planning or thinking a lot.
Pro model are better and straight forward, but it runs out of fuel before you even start prompting. It’s great if you can afford the ultra plan. Pro plans can be optimized using the flash models if you can prompt accordingly.