OpenAI invents a new metric to convince companies their massive AI bills are worth it

Companies have poured enormous amounts of money into artificial intelligence, yet many executives still cannot clearly explain what they are receiving in return.

OpenAI thinks it has the answer. Rather than asking businesses to reduce their AI spending, the company wants them to adopt a new measurement that could make those costs look more reasonable.

It calls the metric… “Useful Intelligence per Dollar.”

The basic idea is that companies should stop focusing on token prices, software licenses, and employee adoption. Instead, they should measure how much useful work an AI system successfully completes for every dollar spent.

That sounds sensible. It is also extremely convenient for a company selling expensive AI models and encouraging businesses to use them for increasingly complicated tasks.

OpenAI argues that the cheapest model is not necessarily the most affordable option. A low-cost model may require repeated prompts, corrections, employee review, and additional computing before it produces something usable.

A more expensive model might complete the same task correctly on its first attempt, supposedly resulting in a lower total cost.

In other words, businesses should not be alarmed by a higher AI bill if OpenAI can persuade them that the final result delivered enough value.

The company says executives should ask four questions when evaluating AI spending.

1. How much useful work did the AI complete?
2. What did each successful task actually cost?
3. How often was the output dependable?
4. Did each AI dollar produce more work as usage increased?

These are reasonable questions, but the answers may be far more subjective than OpenAI suggests.

A customer service department could count resolved support cases. A software company could track AI-generated code that passed testing. A legal department could measure contracts reviewed accurately and on time.

Other outcomes are harder to quantify. How should a business measure whether an AI-assisted meeting produced a better decision? Who decides whether a financial analysis was useful? How much value should be assigned to a presentation that an employee might have created without AI?

OpenAI wants businesses to define what “done” means for each workflow and measure the result inside the system where the work happens.

For finance teams, the company describes employees gathering forecasts, moving data into spreadsheets, reconciling tabs, rebuilding slides, checking calculations, and preparing for review meetings.

OpenAI says ChatGPT Work can perform much of that preparation, allowing employees to concentrate on what changed and what the company should do next.

The sales pitch is obvious. ChatGPT is no longer being positioned as a tool that helps employees write emails or summarize documents. OpenAI wants it embedded deeply inside the daily operations of entire companies.

The deeper that integration becomes, the more difficult and expensive it may be for a business to leave.

OpenAI also says companies should calculate the complete cost of each successful task. That includes model fees, employee time, human review, retries, and any work required to fix errors.

The company recommends adding those expenses together and dividing the total by the number of tasks that reached an acceptable quality level.

This could provide a more realistic figure than cost per token. It could also give companies enormous freedom to manipulate the results.

A business eager to prove its AI strategy is working could define success generously, underestimate review time, and assign inflated values to completed tasks. Executives would then have an impressive metric to present while the actual financial benefits remained unclear.

OpenAI also uses the argument to promote its GPT-5.6 model family.

GPT-5.6 Sol is the company’s flagship model. Terra is intended to balance performance and price, while Luna is marketed as the fastest and least expensive option.

OpenAI suggests businesses could use Luna for routine high-volume work, Terra for tasks requiring more analysis, and Sol when stronger reasoning might reduce retries.

That arrangement also gives OpenAI a product for almost every budget and workload. Rather than deciding whether a task needs AI at all, customers are encouraged to decide which OpenAI model should perform it.

The company claims GPT-5.6 Sol achieved a score of 72.7 percent on the DeepSWE v1.1 software engineering benchmark. OpenAI says that result beat Claude Fable 5 at 69.9 percent while carrying a 36.2 percent lower estimated API cost.

These figures may look persuasive on a chart, but benchmark performance does not guarantee savings inside a real company.

Corporate systems are messy. Data can be incomplete, internal tools may not work together properly, and employees can spend substantial amounts of time checking AI output. A model that performs well during a controlled coding evaluation may behave very differently when deployed across a business with decades of technical debt.

OpenAI says dependability should therefore become another part of the scorecard.

It recommends placing AI results into three categories: ready to use, needs correction, and needs escalation to a person.

This could expose hidden costs. A cheap response that requires an employee to rewrite half of it may be more expensive than a higher-priced answer that works immediately.

The problem is that the company selling the models is also helping define the measurement used to determine whether those models are worth buying.

OpenAI acknowledges that organizations need strict controls before AI systems begin taking actions. Companies must determine what data an AI can access, which systems it can change, and when human approval is required.

Those controls are not minor details. An AI system connected to financial records, customer databases, internal communications, or production code can create serious problems when it makes a mistake.

The cost of a successful task means little if one incorrect action causes a security breach, deletes important data, sends inaccurate information to customers, or creates a legal problem.

OpenAI’s final measurement asks whether AI economics improve as usage grows.

Businesses are told to track successful tasks, total spending, and average cost per completed result. When completed work grows faster than spending while quality remains stable, the company can declare that its AI investment is producing more value.

OpenAI connects this argument to its huge investments in computing infrastructure. Better chips, improved algorithms, smarter model routing, and more efficient data centers could theoretically lower the cost of producing AI output.

Customers may eventually benefit from those improvements. OpenAI benefits immediately when businesses become comfortable measuring success in a way that encourages greater usage.

Useful Intelligence per Dollar is not a useless concept. It is better than celebrating token consumption, purchased licenses, or the number of employees who opened ChatGPT during the month.

Still, companies should be skeptical when an AI vendor introduces the scorecard that will be used to justify buying more AI.

The real question is not how much intelligence a company receives per dollar. It is how much verifiable profit, productivity, or cost reduction the business receives after accounting for every subscription, token, integration, consultant, employee review, mistake, and failed experiment.

Until companies can answer that question with hard numbers, Useful Intelligence per Dollar risks becoming another flattering metric designed to make enormous AI bills look like progress.