Forget Tokens: Why CFOs Need a New AI Scorecard Now

The conversation around AI spending has been stuck on a tired metric: cost per token. For CFOs actually trying to justify nine-figure AI budgets, that number is borderline useless. What they really need to know is the full cost of getting a job done right, not the price of a single word a model spits out.

The real metric, as one industry insider puts it, is “Useful Intelligence per Dollar.” It’s a simple-sounding idea with a complex calculation. You don’t just look at the API bill. You add up employee time spent prompting, the cost of human review, the compute wasted on retries, and the rework when the AI hallucinates. Then you divide that total by the number of tasks that actually met the quality bar. A cheaper model that requires five attempts to draft a contract might end up costing more than a frontier model that nails it in one pass.

This is the logic behind tiered model families like the newly released GPT‑5.6. Its three tiers—Sol, Terra, and Luna—are built for matching capability to task complexity. A customer support bot might fly on the fast, cheap Luna tier, while a financial forecasting workflow that demands deep reasoning is better suited for Sol. The latest benchmarks back this up: GPT‑5.6 Sol hit a new high on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model. It’s not just about being smarter; it’s about being more efficient in producing a successful outcome.

The shift in thinking needs to be cultural, not just technical. The article argues for starting with one workflow and defining what “done” looks like in the real world—a resolved customer issue, a code change that passes tests, a financial forecast ready for review. That’s the output that matters. Tokens only create value when they transform into work people can actually use. As models get more capable and handle longer, multi-step tasks autonomously, the gap between a cheap token and a completed, reliable task will only widen. For businesses, the winners won’t be those who bargain-hunted for the lowest per-token price, but those who best understood the total cost of a correct answer.