Hot take about Claude Opus 5
Opus 5 is a training ground to understand why humans disagree with what it's doing. There's insight there that you can't get without otherwise doing a substandard job.
When a model is, say, 95% good at what it does (by a human standard of completion), there's less reason to explain the refinements needed to make any changes to get it to 100%. "Tweak this, modify that" means very little to an AI system.
When a model is, say, 80% good at what it does (by the same standard), some users will get the shits with it and leave, while many others will work with it to build out the improvements such that the model is more robust at the work that it's doing. Those insights provide Anthropic with valuable training input as to what humans want out of the systems that they are trying to build.