OpenAI’s Cost Playbook Arrives as Agentic AI Bills Keep Climbing

OpenAI cut the price of a million tokens by 97% between GPT-4 and GPT-5.4. Enterprise AI bills climbed anyway. On Tuesday, the company published a five-step playbook telling businesses how to track, evaluate and govern AI spending now that agentic tools burn through far more tokens than a simple chatbot query ever did.

The advice arrives from the same company that profits from that consumption, and its own prescriptions for runaway spend end at OpenAI’s paid capacity tiers. Data from Flexera and Protiviti released in the past few weeks show how unready most enterprises are for either the discipline or the bill.

OpenAI’s Five Steps Start With a Blunt Audit

OpenAI posted the guidance to its own site on July 14, arguing that the shift from simple chat tools to autonomous, multi-step agents means token price is the wrong yardstick for judging whether AI spending is paying off. Cheaper tokens, the company said, do not automatically mean cheaper outcomes once an agent is looping through tool calls and retries to finish a task.

“Without that visibility, a growing bill is hard to interpret,” OpenAI wrote, describing the first and most basic step: knowing who is using AI, which models, how much capacity, and what work it supports. The other four steps move from measurement to money.

Each step builds on the last. OpenAI wants the workflows that survive that funnel to repeat often enough, with clear enough ownership, to justify the investment, a bar meant to separate real productivity gains from expensive experiments.

Cheaper Tokens Still Mean Bigger Bills

The multiplier is mechanical. A chatbot answering a question triggers one inference call. An agent completing a multi-step task calls outside tools, checks its own work and retries when something fails, sometimes 10 to 20 times for a single request.

Gartner concluded in a March analysis that agentic workflows can consume five to 30 times more tokens per task than a standard chatbot query. Goldman Sachs projects global token consumption will multiply 24 times by 2030, reaching 120 quadrillion tokens a month.

At the FinOps X 2026 keynote in San Diego, J.R. Storment, executive director of the FinOps Foundation, said many companies were already three times past their 2026 token budgets, with the year barely half over.

Forrester has a name for the behavior driving that overrun: tokenmaxxing, described in a company blog post as “what you get when you incentivize adoption without governing value.”

Uber offers a plainer example. Chief technology officer Praveen Neppalli Naga told The Information in April that employee use of Anthropic’s Claude Code exhausted Uber’s entire 2026 AI budget within four months. The ride hailing company capped employee AI tool use in June to bring spending back under control.

Two Thirds of IT Teams Are Flying Blind

Flexera, a technology spend and risk intelligence firm, surveyed 512 IT asset management professionals worldwide for its 2026 State of ITAM report. The results describe an industry adopting AI faster than it can track it.

• 59% of respondents said wasted AI spend increased year over year
• 31% reported accurate visibility into their organization’s AI software usage
• 36% said they had a complete picture of their IT assets overall, a decline from prior years

“AI is changing the economics of IT faster than most organizations can adapt,” said Becky Trevino, Flexera’s chief product officer. She described a familiar cycle: fast adoption first, then a scramble to rebuild visibility once the bill arrives.

A Confidence Gap in the C-Suite

Enterprises are not just struggling to see AI spend. They disagree internally about whether it is working. A Protiviti survey of 852 global C-suite executives, conducted with the University of Oxford, found 61% of CIOs and other technology leaders confident that AI is driving revenue growth. Among CEOs and board members, that figure was 30%.

Kim Bozzella, Protiviti’s global CIO and CISO solutions leader, said stronger executive alignment lifts AI confidence across the C-suite. Even ambitious transformation plans stall when “leadership teams aren’t aligned on what success looks like,” she said.

That gap matters for cost control specifically. A CIO watching token usage line by line and a CEO judging AI purely on quarterly revenue are measuring two different things, and neither view alone tells a board whether the spending is working.

Shutterstock Found $250,000 Just by Asking

Shutterstock, the stock imagery and video marketplace, offers a working example of what OpenAI’s playbook looks like in practice. Courtney Totten, the company’s chief technology officer and chief information security officer, described the shift during a FinOps X 2026 keynote.

Understanding your AI costs is no longer optional, it is foundational to business strategy.

Totten renamed her cost management group AI and Cloud FinOps (financial operations, the discipline of tracking cloud and AI spend) and centralized ownership of every AI bill under that team. Output tokens, generated every time a customer searches the site or asks for a generated image, make up roughly 75% of Shutterstock’s total token use.

Centralizing that view paid off almost immediately. Totten’s team found $250,000 in unused vendor commitments about to expire, money that was going out the door regardless until another business unit stepped in to redirect it toward revenue-generating work instead.

Where Does the Discipline Actually Lead?

OpenAI’s five steps end at a purchase decision. Visibility and evaluation identify which workflows earn their keep, governance clears them to expand safely, portfolio funding pays for the buildout, and the last step moves proven winners onto OpenAI’s own paid capacity tiers.

None of that is unusual for enterprise software. Vendors have always built consumption ladders that reward customers for standardizing on their platform. What stands out is the size of the discretion involved: a workflow a governance committee judges strategic gets funded centrally and pushed onto dedicated infrastructure, either Guaranteed Capacity or Scale Tier, or handed to Deployment Company, a partner OpenAI works with to build and manage agents for enterprise clients. Every workflow that graduates through the funnel compounds OpenAI’s own revenue along with it.

Enterprises that skip the earlier steps, the ones Flexera and Protiviti data suggest most still are, arrive at that final purchase decision with neither the visibility nor the internal agreement to negotiate it well. The Tokenomics Foundation, a Linux Foundation program backed by Accenture, Google Cloud, IBM, JPMorgan Chase, Microsoft, Oracle, Salesforce and SAP among others, is trying to standardize AI billing before that gap widens further. Its successor event to FinOps X, called Tokenomicon, is set to debut in San Diego from June 7 to 10, 2027.

Frequently Asked Questions

What is tokenmaxxing in AI spending?

Tokenmaxxing describes internal incentives, like leaderboards or usage quotas, that reward employees or teams for consuming more AI tokens rather than for the value that usage creates. Forrester coined the term to explain why some companies saw bills spike even as actual output stayed flat, a pattern distinct from healthy adoption because nobody was measuring whether the extra usage did anything useful.

What is the Tokenomics Foundation?

The Tokenomics Foundation is a Linux Foundation program announced at FinOps X 2026 to build a shared billing schema and benchmarks for AI token spending, similar to the role the FOCUS specification played in standardizing public cloud bills. Backers include Booking.com, KPMG and ServiceNow alongside firms already managing enterprise AI spend, with the goal of making token invoices comparable across providers.

Why do AI agents use so many more tokens than a chatbot reply?

A single agentic task can involve planning steps, calls to outside tools or databases, verification of its own output, and repeated retries when a step fails, each consuming additional tokens beyond the original request. Output tokens also cost several times more per token than input tokens, which compounds the effect further as an agent reasons through a problem.

How can IT teams start tracking agentic AI costs today?

Most enterprise AI platforms, including OpenAI’s Admin Console, already break down usage by user, team, model and product, so the first practical step is turning that reporting on and reviewing it weekly rather than waiting for the invoice. Pairing usage data with a defined outcome, like a resolved ticket or a passed code review, shows whether rising usage reflects growth or waste.

What is FinOps and why is it now a boardroom topic?

FinOps, short for financial operations, is the discipline of managing and forecasting cloud and technology spend across engineering and finance teams together. It grew up around public cloud bills, but token based AI pricing has pushed it into the boardroom because agentic AI spend can move by tens of thousands of dollars in a single week without any new headcount or contract change.