A cost footnote for the “steer a long run” part: on Opus 5.5 thinking can’t be switched off (disabled thinking or a manual budget returns a 400), so effort, medium by default, is the only lever, and reasoning bills at the $20/M output rate. Effort is part of the prompt-cache key along with the model, so changing either mid-run makes the next request re-read the whole context without cache hits. That costs more than usual here, because Opus 5.5 reads cache at $0.20/M, 5% of its $4 input price rather than the usual 10%: re-reading a 200K-token context is about $0.04 from cache, $0.80 as fresh input, and $1.00 if it has to be written back to the 5-minute cache at $5/M. On Pro or Max the same thing shows up as plan allowance rather than dollars, and Anthropic doesn’t publish the token-to-limit conversion, so the practical advice is to set model and effort at the start of a session and leave them alone.