> I remember when they would always hallucinate APIs that wasn't there or make up fields that didn't exist.. those problems are virtually solved now.

I love how people have said this for over a year now, and no matter how often I try it, it's still just as broken.

Try it with any task that isn't in the training data. Say, a custom protocol where you need to cross-reference multiple implementations and the specs to really get it, and with no answers on stackoverflow or medium.

At first it's hilarious, but after a while it just gets exhausting. For all these real-world tasks you need to put so much effort in that it's much easier to just write the code yourself, even with the latest (current Gemini) and greatest (Opus 4.6/4.7) models.

If the parts for your project don't already exist, AI can't help you either. And if they do exist, why spend money on AI if the GitHub search is free and you can just fork and modify what you need?