The tokenmaxxing era ends as companies struggle to measure AI effectiveness amid spiraling costs

The AI tokenmaxxing party is crashing over spiraling costs -- leaked consulting firm audio suggests no one is sure how to measure AI effectiveness

"Leadership [...] are still asking the question of whether they're getting value from what we're spending." The era of AI tokenmaxxing may be well and truly over. Alongside stories of Amazon cutting its AI leaderboard and an unknown company blowing through $500 million worth of tokens in one month, leaked audio has emerged from consulting firm Accenture as it tries to figure out how to rein in rampant token spend at client companies, 404Media reports. In leaked audio, Accenture acknowledges that certain trivial tasks being offloaded to AI are causing massive token overspend, especially when agentic AI is part of the mix. The staff in the meeting clearly recognizes that not only is AI spend growing out of control at companies heavily adopting the technology, but that there is very little way to predict how much any tasks would cost, or whether there is real value in using AI to complete them. Accenture has previously been incredibly bullish on AI, even encouraging employees to use it so much that if they didn't, they risked missing out on promotions. But that seems like a policy destined for the AI history books, as Accenture is now clearly aware that it's overspending on AI, and many of its clients are too. From tokenmaxxing, to token hoarding For much of the past year, many companies have charged full speed into an AI-heavy business strategy. Amazon had an AI leaderboard, and Nvidia's CEO Jensen Huang said he'd be alarmed if engineers weren't spending at least 50% of their annual salary on AI tokens. Anecdotally, I know a number of software developers and data engineers who have been encouraged to use AI as much as they can. They have token limits, but they have been encouraged to use all of them and find new ways to do it, too. This is leading to runaway token spending, something Accenture is seeing in its client data. Accenture's agentic AI strategy lead, Justive Kwak, was quoted in the audio saying: "What we're seeing right now is just rapid escalation in AI token spend [...] as companies start to scale AI, moving from like simple chatbots into use cases that feature agentic workflows and automation and then enterprise-wide deployment of some of these tools like Copilot, Claude Code, and Codex." This isn't something that will be contained to just a few firms, either, he said. "It's really not a niche problem. It is a problem that every enterprise will face if they are bullish on AI, if they haven't already," he said, adding that token spend was increasing, "exponentially, as more and more people are starting to use AI." But that may be starting to change. Amazon canned its AI leaderboard - it's rumored to be the mystery company with a half-billion dollar AI spend in one month - Uber is capping AI use to cut costs, and Axios reported at the end of May that a number of CEOs and companies were switching to more affordable models, and more closely monitoring employee usage. Some software developers I know have been using the "caveman" trick to reduce token spend. Even OpenAI CEO Sam Altman said that he was aware token costs were becoming a huge concern for people. This all comes in the aftermath of the move by many of the major AI providers to token-based billing. Where previously subscriptions offered very favorable rates for AI use, suddenly companies were having to pay for the tokens they input, and the tokens the AI output - even when it was verbose, or made mistakes, or required follow-up correction. As the Accenture call shows, it's making even some of the most AI-bullish organizations question their usage, because measuring the spend and the return on that investment is proving all but impossible. As Kwak said in the leaked audio, "Leadership, especially at the CFO, COO, and CIO level, are still asking the question of whether they're getting value from what we're spending on in the context of AI." How do you measure return on investment? Although large language models are proving to be extremely useful in niche cases, their effectiveness at a broader range of tasks is more nebulous. Especially when it comes to financing it. When managers and executives look at AI budgeting and a return on that investment, it's hard to square away the numbers. When you can't know how many tokens a task will take to complete, or whether the task will be completed effectively on the first, second, or third attempt; when you can't completely control the length of the output, or know whether that output will be wrong, or a lie, or just a random hallucination, how do you measure return on the investment in that tool? "We're hitting this inflection point where AI is becoming material to the cost structure; spend is becoming very unpredictable," Accenture's Kwak said during the meeting. Although the overall bill of AI costs is visible, he suggested, finding the specific value attributed to that token spend was not. This seems to have created a culture of task hierarchy within Accenture, where some tasks are deemed more worthy of AI token use than others. When Kwak positioned himself to show some slides during the meeting, Accenture's client group lead, Stuary Henderson, joked that he hoped Kwak didn't use AI to convert a PDF into images and then markdown files. "I'm learning that's one of the big token chewers," he said. "Turning PDFs into markdown: is that right?" Kwak agreed that Accenture data did show some tasks being completed using AI that didn't really need it, and were using unnecessary tokens because of it. Much of that problem, he suggested, was down to non-technical staff overusing it. "We're seeing from some of the data internally at least that it's actually not our engineers that are driving the token consumption. It's a lot of the non-engineers that are doing some of those behaviors." Now that Accenture has encouraged heavy AI adoption among its clients, it finds itself in the bizarre position of having to discourage it or at least encourage more studious use of it. It now sees its next opportunity as a way to advise clients on how to "think about token economics." It's working on a tool called "Token IQ" to help advise clients, according to the call, but hasn't made any announcement so far. What's clear from the Accenture leak and actions of some of the major tech companies, which have previously been so bullish on AI use, is that the finances of mass AI adoption at the per-token scale don't line up. Without a clear way to measure the return on AI investment, we may find even the most tokenmaxxing companies look to restrict access and spend through the rest of 2026 as they re-address AI strategy.

OpenAI and Anthropic face new AI reality as companies shift from tokenmaxxing to efficiency

"Some of their largest enterprise customers may start limiting their out-of-control token spend," said D.A. Davidson analyst Gil Luria, regarding concerns around OpenAI and Anthropic. Flo Crivello's expenses were out of whack, and there was only one way to get them under control. Earlier this month, the 34-year-old CEO of AI startup Lindy switched his company off of Anthropic's Claude models, moving 100% of its traffic to DeepSeek, a Chinese company that makes cheaper, open-weight alternatives. "We did it, and you could see that cost curve go down, like, crash to the ground," Crivello said in an interview from his company's San Francisco headquarters. He said the decision will save Lindy millions of dollars within months, though he still expects the roughly 25-person company to spend more on AI than payroll. "It's a matter of survival for the business," Crivello said. "That's all it is." Crivello, who previously spent almost five years at Uber, is among a growing crop of founders and executives across the U.S. trying to rein in artificial intelligence spending. Bills for AI have ballooned - sometimes into the billions of dollars - since OpenAI first captivated Wall Street with its ChatGPT chatbot in 2022, kickstarting a rush by businesses to deploy the technology across areas like customer support, marketing and finance. In particular, costs ramped up in the realm of AI-assisted coding, as developers pumped tokens into the creation of new tools and services that previously would have required teams of coders. That led to the era of so-called tokenmaxxing and AI leaderboards, where employers have incentivized developers to use as much AI as possible without worrying about the results. The crackdown is underway. Uber said this month it had implemented a series of spending tiers on some AI tools, starting at a base level of $1,500 per month, though employees could request access to higher levels. In April, Uber CTO Praveen Neppalli Naga revealed to The Information that the ride-sharing company blew through its entire annual AI budget in just four months. OpenAI and Anthropic have been the principal beneficiaries of the spend-at-all-cost mentality, which has fueled their exponential growth rates and pushed both of the AI model leaders to valuations approaching $1 trillion. Now, as they gear up for potentially historic IPOs -- both filed confidentially in early June -- the mood around AI is shifting, and business leaders like Crivello are no longer willing to throw money at Anthropic or OpenAI without a clear picture of a return on their investment. "Current growth rates for Anthropic and OpenAI are the fastest they will ever be, which is mostly a matter of basic math," Gil Luria, an equity analyst covering tech companies at D.A. Davidson, told CNBC. "That is a good reason to go public now, as is the concern that some of their largest enterprise customers may start limiting their out-of-control token spend." Anthropic last reported a $47 billion annualized run rate in May, up from the roughly $10 billion in revenue it recorded for all of last year. OpenAI's run rate was pacing closer to $25 billion earlier this year, according to reports, up from the $13.1 billion in revenue it generated in 2025. Listing soon, while the numbers are still dazzling, could be strategic. "There has to be some period of time in the future where there's some rationalizing of spend by companies, and that may be a blip ahead for Anthropic and OpenAI," Luria said in an interview. "That creates some sense of urgency to go public before we see that." Anthropic declined to comment for this story. OpenAI didn't respond to a request for comment. Crivello said he's a big fan of Anthropic, but his company had been dealing with "unsustainable" AI costs for a long time. Lindy was built around the idea that the cost of tokens, or the units of data that are processed and generated by AI models, would decrease dramatically over time, Crivello said. That proved true for a while, but leading model developers, including Anthropic and OpenAI, have been slower to slash prices in recent months. Crivello said he'd be open to switching Lindy back to Claude models if the prices come down. "I hope that they cut the costs again at some point but, until then, we've got options," he said. Jeff Henry, president of consulting at Highspring, said some of his firm's clients are pulling back until they "can really start to prove an ROI," and others are still waiting another 12 to 18 months before making any big spending decisions. "Everybody is experiencing the same spend crunch on AI," he said. However, there are still countless mid-sized companies that haven't even started experimenting with AI yet, he said. "AI is not going away," Henry said. "There's no way that toothpaste ever goes back in the tube." Darren Kimura, CEO of enterprise AI company AISquared, said one area where AI spending is "absolutely" hitting a peak is in use of state-of-the-art models, also known as frontier models, for simple tasks that can be accomplished with cheaper alternatives. Some companies are turning to what's called model routing, which matches the appropriate task to the appropriate model. It's a technique so new that, according to Glean CEO Arvind Jain, roughly 95% of enterprise AI usage is still running on frontier models. Kimura said that approach will be "untenable" for most companies in the long run. D.A. Davidson's Luria said pricing in the market is still at an "unsophisticated" stage, but both OpenAI and Anthropic have been trying to adjust to an increasingly budget-conscious environment. OpenAI launched analytics and updated controls for enterprises earlier this month, allowing administrators to break down credit spend across the workplace, set usage limits and give employees visibility into their available budgets. Anthropic rolled out a series of controls in August that allow customers to provision users, view analytics and set spending limits at the organization and individual level. Finance departments are paying close attention after getting hit with surprisingly large AI bills, said Eric Glyman, co-CEO of expense management startup Ramp. "Most CFOs not only didn't plan for this in their annual plans -- the steep growth -- but don't have great tools to manage this," Glyman said in an interview. "Suddenly you have this third pillar that has showed up, which is spending through tokens and intelligence. It's not a clean area of spend." As companies become more price sensitive to AI, OpenAI and Anthropic have to contend with deep-pocketed competitors that are aiming to develop lower-cost models. Microsoft, which has poured more than $13 billion into OpenAI as much as $5 billion in Anthropic, unveiled a suite of new low-cost models earlier this month. The company has also emphasized that its AI coding product, GitHub Copilot, will route users to the most appropriate model for a task. In a June essay, Microsoft CEO Satya Nadella said the industry needs to avoid concentrating power in a handful of large providers. "The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see," Nadella wrote. "If all the value is accrued by only a few models, the political economy will simply not tolerate it." Amazon and Google are also ramping up their investments in models for business users. Peter DeSantis, Amazon's top AI executive, told CNBC this month that he hopes the company will be able to compete with OpenAI and Anthropic's frontier models in the "coming year." Like Microsoft, Amazon is an investor in both of those companies. DeSantis said in February that Amazon will rely on its in-house chips to develop models at a less expensive rate than its rivals. "AI has a cost problem," he told The Wall Street Journal in an interview. "If we ultimately want AI to transform everything, the costs have to be different." Google made a concerted effort to highlight affordable AI offerings at its annual developer conference last month. The company showcased Gemini 3.5 Flash, a lighter-weight addition to its model suite that's available at half, or in some cases close to one-third, the price of comparable frontier models, according to CEO Sundar Pichai. "Microsoft and Google have the infrastructure and capability - the entire stack - where they can come in and stiff-arm both OpenAI and Anthropic," PitchBook analyst Harrison Rolfes said in an interview. "They're probably waiting on the sidelines for them to battle it out, see where they're not doing well." As for going public, neither of the big model companies have provided an exact timeframe for their prospective debuts. The New York Times reported on Thursday, citing people involved in the deliberations, that OpenAI is leaning toward holding off until next year. Pressure to go public may revolve around the need for capital. With Anthropic and OpenAI increasingly competing against their biggest financial backers, the IPO market may be the best avenue for new money, especially as their capital needs have become too great for most venture and private equity firms. "A lot of the traditional pockets of capital are drying up," said Dharmesh Thakker, a general partner at Battery Ventures. "All the institutional investors who can invest in these companies have already taken their pound of flesh." Choose CNBC as your preferred source on Google and never miss a moment from the most trusted name in business news.