OpenAI’s CFO Sarah Friar presented a new framework for evaluating corporate investments in artificial intelligence on July 17, 2026. Instead of token prices or pure usage numbers, four questions are intended to clarify whether the use of AI is economically worthwhile. The initiative appears in a blog post by OpenAI and responds to growing skepticism among many companies regarding their AI expenditures.
Four Questions Replace the Focus on Pure Token Prices
As OpenAI states in a blog post, the metric “Useful Intelligence per Dollar” is meant to replace traditional software metrics such as user numbers or contract renewals. The first question concerns the work done itself: Does a system actually resolve customer issues, deliver code, or review contracts, rather than just answering queries? The second question targets the full costs per successful task – including rework, human oversight, and repeated attempts, not just the pure model price.
Thirdly, it is about reliability: Results should be directly usable without correction or escalation; otherwise, costs shift invisibly onto employees. The fourth question examines whether the value of the completed work grows faster with increased usage than the costs incurred for it. Friar sums up the fundamental problem: The crucial factor is whether the economic benefit of AI work increases faster than its production costs. Unlike traditional software, the success of AI systems cannot be solely assessed by subscription numbers or active users, as a single task may involve multiple model calls and control steps.
New Model GPT-5.6 Accompanies the Initiative
In the same blog post, OpenAI links the new metric to the introduction of GPT-5.6 in the three tiers Sol, Terra, and Luna, which cater to different budgets and task profiles. The Sol version reportedly achieves top performance in programming tasks while requiring 54 percent fewer output tokens than comparable models – a claim that is not independently verifiable. Terra and Luna are said by OpenAI to offer more affordable alternatives for high-volume, lower-complexity tasks.
The timing suggests that the new measurement method is also intended to justify higher prices for more powerful models, as it shifts the focus from token quantity to value. At the same time, OpenAI is now investing up to 500 billion US dollars over four years in data centers through the infrastructure initiative Stargate, significantly more than the original target of 100 billion dollars. These sums increase the pressure to present tangible economic results to investors and customers instead of mere usage growth figures. For corporate clients, this likely means that pricing models will increasingly orient themselves towards verifiable results rather than pure computing power.
Customers Respond Cautiously to Rising AI Expenditures
According to Axios, which first reported on the initiative, a corporate client recently accidentally received a bill equivalent to half a billion dollars for a single month. OpenAI CEO Sam Altman now cites costs as the second biggest concern for customers, right after difficulties in the practical implementation of AI systems in corporate everyday life. Already in January, Friar reportedly pointed out a mismatch between the capabilities of AI systems and the actual business value captured, according to Fortune.
According to Fortune, the new framework also has an economic self-interest: An evaluation based on work results rather than token price tends to favor more expensive, powerful models like those from OpenAI itself. Many companies are currently outsourcing routine tasks to cheaper competing models instead of consistently opting for more expensive systems. The initiative thus encounters a clientele that is actively seeking ways to reduce AI expenditures. The new standard remains comparably limited as long as each company sets its own criteria for when a task is considered successfully completed.
It will be crucial whether the four metrics establish themselves as an industry standard or whether, in the end, each provider proposes its own measure for the value of its models. For financial decision-makers, the practical hurdle remains to accurately capture costs for rework and human oversight in existing accounting systems.
