The most common money mistake in AI isn't picking a bad model — it's picking a great one for a task that didn't need it. Summaries, reformatting, extraction, classification: a model that costs 20x less often does these indistinguishably well.
The operator habit: before wiring any repeated task into a workflow, race a frontier model against a budget model on that exact task. If you can't tell the outputs apart, the cheap model wins — permanently, at scale.
Run the seed: a routine extraction task. Compare a frontier model against DeepSeek or Gemini Flash. Be honest about whether the expensive output is actually better, or just more expensive.