New AI Trick Cuts Chatbot Costs 74% — With the Same Answers

When you ask an AI assistant to do something — book travel, edit a spreadsheet, answer a policy question — it often pulls a ready-made instruction sheet from a library. Researchers call these "Skills." The trouble is that the AI loads them blindly. A Skill can look relevant but be unnecessary, expensive, or even harmful to the task at hand. Shuang Guo's new paper introduces SkillApt, a small gatekeeper that sits between "should I grab this Skill?" and "should I actually use it?"

SkillApt works by testing. It runs a task twice in similar past situations — once with the Skill, once without — and compares which came out better. Then it makes a simple load-or-skip call. Think of a chef tasting a spice before dumping it in the pot. On a standard test called SRA-Bench, SkillApt got exactly the same accuracy as the usual method (0.838 versus 0.838) but used Skills only 31.5% of the time instead of 100%, and burned 74.3% fewer tokens — the word-pieces AI reads and you pay for.

For you, that translates into cheaper AI tools and faster replies. Companies running AI agents pay per token, so cutting usage by three-quarters while keeping quality is real money — potentially tens of thousands of dollars a month for busy operations. It also means fewer weird failures, since loading the wrong Skill can derail an otherwise fine answer. The paper's bigger idea is that picking a Skill and using it should be two separate decisions, not one.

The catch: this is a preprint, meaning it hasn't been checked by other scientists yet. It was tested on one benchmark, and the authors found the benefit varies depending on which AI model you start with. Some models learn the load-or-skip boundary better than others. So don't expect a universal 74% savings — expect a useful direction that AI builders will likely copy.