It's the cost per task. Sometimes it happens that the higher-effort model solves the problem in fewer tokens, thus making the overall price cheaper. It even happens with models (from the same or a different provider) that are priced differently.