OpenAI proposes a CFO scorecard for measuring useful AI work per dollar

OpenAI is trying to shift the enterprise AI conversation away from raw model prices and toward a broader measure of whether AI systems are completing valuable work. In a July 17 company post, CFO Sarah Friar argued that buyers should judge AI spending by "Useful Intelligence per Dollar," a scorecard built around work completed, cost per successful task, dependability, and whether the economics improve as usage scales.

The framing matters because many businesses still compare AI tools through visible inputs such as seats, tokens, or subscription prices. OpenAI says those numbers can miss the real cost of an outcome. A cheaper model may need more attempts, more review, or more rework, while a more capable model may finish a complex task in one pass. The company is encouraging customers to define a workflow, decide what a completed result means, and measure whether AI reduces the total effort needed to reach that result.

Why OpenAI is emphasizing outcomes

The post connects the scorecard directly to enterprise adoption of agentic systems. OpenAI points to examples such as support tickets resolved, code changes shipped, contracts reviewed, financial forecasts prepared, and decisions improved by timely context. Those are not benchmark-only measures; they are operational results that can be checked inside the systems where work already happens.

OpenAI also used the post to reinforce its GPT-5.6 positioning. It said the GPT-5.6 family has three tiers: Sol as the flagship model, Terra as a balance of performance and cost, and Luna as the fastest and most affordable option. The company claims GPT-5.6 Sol with maximum reasoning reached 72.7% on DeepSWE v1.1, above Claude Fable 5 at 69.9%, with a 36.2% lower estimated API cost. It also said GPT-5.6 Sol set a new high on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model.

Those figures are OpenAI's claims, but the broader message is aimed at procurement and operations teams: model selection should be tied to successful task economics, not just sticker price. The post says organizations should track whether each task was ready to use, needed correction, or required escalation to a person. That kind of accounting could become more important as AI moves from drafting text to using tools, touching company data, and carrying out longer workflows.

For customers, the practical takeaway is to start with a narrow workflow and measure it carefully. For vendors, OpenAI is signaling that future competition will be about dependable completed work, efficient inference, routing across model tiers, and governance controls that make higher-value automation acceptable inside enterprises.

Sources

• openai.com

Cover image: Robert Scoble, source, licensed under BY.