Okay, maybe I was just unlucky with Astra. I ran Sol 6.2 high, it comes out much better. He completed several tasks himself and reported clearly and comprehensibly. My previous Sol 6.1 on similar missions often went into some endless changes and didn't always get me right.
And I don't get arguments about "you just set it up wrong." " Because I set all the gears and models in about the same way, but with carbon twins I have no problem understanding both ways. And he clearly understands what I want from him, and I understand his messages.
So we won't bury the new release of OpenAI prematurely. I keep watching.