Anecdotes and conspiracies and mass hysteria abound imo. And probably a bit of astroturfing from competitors. Models are not deterministic, nor are your projects. I work on code constantly for years - and across every anthropic and openai model people claim the same thing - yet I've never experienced this degredation, not even one hint of it. I would be acutely aware of it.

Variable token usage? Sure. To be expected in non-deterministic systems like these.

Distinct increase in errors and issues? Never. Not within a single model's lifespan have I ever experienced this.

And this whole notion of the A/B testing of performance/usage is completely ridiculous. There is zero evidence it has anything to do with performance. There are a trillion things to A/B test besides things that hamstring groups of users, so this is just being used to confirm the bias of people already caught up in these "theories."

I await any realistic indication.