Price wars for AI models have shifted from cutting costs to “using less compute.”
Opus 5 still charges $5 per million for input and $25 per million for output. According to Anthropic’s self-tests, at the highest effort on CursorBench, it trails the peak of Fable 5 by only 0.5%, yet the cost per task is cut in half. Axios also notes that this is the fourth Claude 5 variant in under two months.
The real innovation isn’t the half-price tag—it’s the effort switch: the same model starts allocating compute based on the task. In the future, enterprises shouldn’t just compare token unit prices; just track these three things—how many tool calls were used to complete a task, how many rounds of rework it took, and how much human intervention was required.
Opus 5 still charges $5 per million for input and $25 per million for output. According to Anthropic’s self-tests, at the highest effort on CursorBench, it trails the peak of Fable 5 by only 0.5%, yet the cost per task is cut in half. Axios also notes that this is the fourth Claude 5 variant in under two months.
The real innovation isn’t the half-price tag—it’s the effort switch: the same model starts allocating compute based on the task. In the future, enterprises shouldn’t just compare token unit prices; just track these three things—how many tool calls were used to complete a task, how many rounds of rework it took, and how much human intervention was required.