Definition
Useful intelligence per dollar is a proposed AI value scorecard connecting useful work, cost per successful task, dependability, and value at scale. It asks whether an AI system produces work people can use and whether that value grows faster than the full cost required to produce it.
Who proposed it
OpenAI CFO Sarah Friar introduced the scorecard in a July 2026 essay. The source matters: OpenAI has a commercial interest in moving buyer attention away from unit token price and toward the value of more inference. That does not make the framework useless, but it means the value claims should be measured by the buyer rather than accepted from a model provider.
Why it is contested
The phrase combines quantities that have no universal unit. Different tasks produce different kinds of value, and "intelligence" is not directly measurable in dollars. A provider can measure benchmark performance and inference cost, but it cannot supply a buyer's task mix, dependability threshold, risk, or business value. This glossary therefore treats the proposal as a scorecard. An operational use must define the accepted outcome and count model, tool, infrastructure, retry, review, correction, and incident costs.
Distinguish it from nearby terms
Token efficiency measures accepted output against token consumption. Cost per accepted durable outcome includes broader costs and a durability window. Useful intelligence per dollar adds the business-value question and therefore requires local valuation, task mix, and risk assumptions.
Check your understanding
A provider says its expensive model delivers more useful intelligence per dollar. What evidence would you need? Run your task distribution, apply your acceptance and dependability rules, count the full cost, and value the accepted result. Provider benchmarks cannot supply your denominator or your business value.