What's this about?
A correction from Anthropic's Thariq: the improvement where changing effort levels no longer breaks the prompt cache is live on the API — but not yet in Claude Code. It did not ship with the model launch, contrary to what many assumed.
What exactly is the prompt cache?
Providers cache a conversation's shared prefix so repeated requests don't re-pay to process it — the main cost lever in long agent sessions. Historically, toggling effort invalidated the cache, making mid-session effort changes expensive.
Why is this worth noting?
Mostly as a correction with real billing consequences for heavy users who started toggling effort freely — and as a small model of good practice: a lab engineer publicly correcting the record within a day.