Case 001: RTK on Terminal-Bench
RTK advertises up to 90% less bash output. In one Terminal-Bench run against no tool, cost per suite pass came in 15.2% higher.
Research
Guilty Until Proven — grading agent trajectories
We replay long real sessions and grade each one against an uncompressed control. Quality holds with 62% fewer tokens and 45% lower cost.
The anatomy of a token bill
Where agents spend money, and how to minmax it. A real coding session, priced cache-aware, comes in 72% under baseline.