When it Comes to LLM Use, Expertise Can Only Take You So Far
In Myth 2, the perception that higher-cost models are more effective at all tasks was debunked. Expert-level users can guide lower-cost "generic" LLMs effectively. But the benefits do not extend to cost management.
Anthropic's analysis of Claude Code usage revealed that expert users generate more output and actions per prompt.
According to Anthropic: "In typical novice sessions, each prompt sets off about five Claude actions and roughly 600 words of output, while expert sessions set off action chains more than twice as long (12 actions) carrying five times the output (3,200 words)."
Depending on the model, the 'expert effect' can drive costs up significantly. Anthropic found that each level of expertise generates approximately 9% more actions and 13% more text output per prompt.
The efficiency gain for expert users is in task completion rate rather than token efficiency. Experts complete more of what they ask for, but they also drive LLMs to deliver more.
This is a good outcome, but model selection also plays a role. If experts can get excellent outcomes from "generic" and "brand name" models, then selecting the most cost-efficient LLM needs to be a priority.
