UCLA study: About 18% of tokens generated by AI agents are wasted

A paper from UCLA and Northeastern University finds that approximately 18% of output tokens generated by AI agents are removable waste; removing them has almost no impact on task success rates. The study covers six cutting-edge models and four benchmarks. In Kimi-K3 inference, eliminating redundant and filler content reduced output tokens by 41% on SWE-bench Verified, with success rates moving from 89.0% to 89.5%. Authors suggest identifying waste sources based on tasks and harnesses rather than simply limiting token counts.