Builders seeking to curb the price of AI-powered coding instruments have more and more turned to the “Caveman” prompting type, which instructs coding assistants to speak in blunt, telegraphic language and keep away from conversational padding. The idea is easy: fewer phrases imply fewer tokens, translating into decrease inference prices for organizations deploying AI brokers at scale.
A brand new check from IDE maker JetBrains confirms that terse prompting kinds such because the viral open-source Caveman project can scale back token utilization with out hurting coding efficiency. Nevertheless, the corporate discovered that the financial savings had been far smaller than supporters declare.
JetBrains used the Harbor open-source analysis framework and duties from SkillsBench for its check, and located that the Caveman approach diminished utilization of output tokens by about 8.5%, far under its claimed 65%.