The issue with Claude are twofold:

1. Poor long context performance compared to GPT5.1, so Claude gets confused about things when it has to do exploration in the middle of a task.

2. Claude is very completion driven, and very opinionated, so if your codebase has its own opinions Claude will fight you, and if there are things that are hard to get right, rather than come back and ask for advice, Claude will try to stub/mock it ("let's try a simpler solution...") which would be fine, except that it'll report that it completed the task as written.

What have you done to make Claude stronger on brownfields work? This is very interesting to me.