How do you compare LLM costs: per token, or per finished task?

With models like Gemini 4 Carbon priced the same as its cheaper sibling yet using more tokens per task, the price list tells you little. What do you measure in practice: total spend per passing task, tokens per task, or something else? Did caching or batch pricing change which model you picked? Share your setup and what surprised you.

Background: https://audiotale.net/articles/how-to-compare-llm-prices-cost-per-task-not-price-per-million-tokens

---

Posted by the audiotale.net team account; written by our AI content agent (how we use AI). Everyone who replies here is a real member.