Anthropic's recent post https://anthropic.com/engineering/april-23-postmortem seems at odds with this blog post. Am I misinterpreting it, or does it indeed say they configured Claude to do less 'thinking' in order to address a performance issue?
My problem isn’t that the models aren’t good. I’ve got that down and opus has been so great since last fall.
My problem is that ever since about a month ago, the token usage has gotten out of control and I’m hitting $200/mo max weekly limits sooner when I used the $100/month max for months without ever hitting a weekly limit.
Wish this post was more technical, e.g., how is the author achieving the things they advise others to do. As is, the post can't seem to decide what it wants to be: is it a rant, is it a tech post, is it a brag, or something else? Are people embarrasing themselves when they're simultaneously affected by three bugs in Claude Code? How did the author avoid being affected by them?
He gave several real examples. The idea is that you have to audit your own agent file and other harness settings. He can't provide an example from your file and one from his doesn't help make the point any better.
I too would have liked to have seen some meat on the bones. I share OP’s frustration but their lack of tech steps left me wanting. Would love to see real examples.