OpenAI published guidance on July 14 telling corporate buyers to retire the metric the company spent two years pushing them toward: the price per token. The new document, aimed at chief information officers managing AI budgets, argues that a shrinking per-query bill says nothing about whether the work being done with it is worth paying for. The company making that argument is also the vendor whose own pricing collapse it cites as evidence.
OpenAI says its API price per million tokens dropped 97 percent between GPT-4 and GPT-5.4. Its newer GPT-5.6 model posts a higher score on the Artificial Analysis Coding Agent Index while using 54 percent fewer output tokens and finishing tasks 57 percent faster. Those figures come from OpenAI’s own materials. The guidance does not cite an outside lab replicating them.
That absence matters given what the document is asking readers to do next. OpenAI wants enterprises to stop shopping on sticker price and start tracking cost per accepted outcome, meaning a support ticket actually resolved or a code change that actually passes review. That standard requires trusting the vendor’s own definition of “accepted,” in a report the vendor wrote.
The guidance also functions as a product catalog. ChatGPT Work, an admin console with usage analytics and spend controls, appears under the recommendation to sharpen visibility into who is using AI and how. Deployment Company, a program that puts OpenAI engineers inside a client’s build process, appears under the recommendation to fund workflows that compound over time. A company advising customers on how to measure return on AI spend has an obvious stake in tools that make its own usage look efficient by that measure.
Stripped of the sales layer, the five practices are reasonable operating advice. OpenAI recommends enterprises: get a plain view of who uses which model and product before treating a rising bill as either waste or growth; judge models by cost per completed task rather than token price alone, saving frontier intelligence for ambiguous or high-stakes work; put governance, meaning defined tool access and human approval steps, in front of any workflow before it scales; fund AI like an investment portfolio, with broad everyday access, function-specific projects, and a small number of strategic bets tied to proprietary data; and match purchased capacity, through options like Guaranteed Capacity or Batch API, to demand that has already been proven rather than projected.
The framing marks a break from what practitioners call tokenmaxxing: treating the lowest per-query price as the main lever on an AI budget. OpenAI’s own numbers suggest that lever has little further to give after a 97 percent price collapse across roughly two model generations. The company is proposing a new one instead, useful work per dollar, and it happens to be the metric on which its newest model claims an advantage.
For CIOs, the practical shift is in what the next budget review measures. A cost report broken out by tokens and model name will not answer what an executive actually wants to know, which is what the spend bought. Enterprises renewing AI contracts this quarter should build their own cost-per-accepted-outcome tracking, using their own acceptance criteria, before adopting a vendor’s efficiency claims as the basis for a multi-year deal.
OpenAI published this enterprise AI investment guidance on its website on July 14, 2026.
