I'd hate to be that guy, but Opus not a very smart model when the effort is set to anything below high. I think, given the feedback from the community, this would be an obvious signal. However, moving the effort to anything beyond medium is a huge token burn. These issues didn't exist, or at least not this persistent, before the last 2 weeks. I, and perhaps a million or so other developers, would ask you to reconsider this thinking. I understand you need to run a business, but so do we, and Claude Opus is genius with a drinking problem, and you never really know upfront if it's drunk or not, but it's generally quite clear after a few minutes.

Other models, such as K2, GLM-5.1, and "the other one" seem to far less drunk than your approach, and you're losing fans quickly if you keep making these kind of changes to the tools or models.