Why it matters
Same model, same tasks, about a fifth off the bill on two of them. The cut came from the harness: Claude was given 3 tool definitions, not 16, so the opening request shrank from about 38,000 tokens to 22,000, and that base rides along on every later request.
Tokenmaxxing read
The usual cost fix is a cheaper model. This one is fewer round trips: the feature task took 6 to 8 requests, down from 10 to 13, and cache reads more than halved, to 211,000–274,000 tokens from as many as 678,000. Count requests per task before you trade down on model quality.
Source takeaway
Mind the limits. The mode is experimental, needs an environment variable, runs solely in a live terminal session, and switches off computer use, Claude in Chrome and Remote Control. The test covered one model on tasks that ran under two minutes.



