Tokenmaxxing is facing a confidence crisis. Just last month, news broke that Microsoft would be cutting most of its Claude Code subscriptions amid financial concerns.

Similarly, Uber’s COO noted that the company has burned through its entire AI budget in just four months, while Axios shared the story of an AI consultant who claimed a client spent $500 million in a month because no one put usage limits on Claude licenses for employees.

More and more companies are sharing cautionary tales about what happens when tokens are used as a productivity metric. Pushing developers to consume as many tokens as possible can encourage AI adoption , but it also opens the door to catastrophic expenditure.

But is tokenmaxxing over as many suggest? While startups like Cleo have seen results from the approach, many other companies are moving away from high token consumption as a strategy. Now there seems to be a shift in the industry away from high token consumption toward “valuemaxxing.”

Tokenmaxxing broke onto the scene as a trend in early 2026, as tools like Claude Code and Codex became increasingly powerful. The trend involved encouraging software developers to consume as many AI tokens as possible to drive innovation. The practice quickly made token consumption an imperfect productivity metric.

It’s worth noting that pushing high token consumption wasn’t a fringe trend embraced by a handful of startups, but a practice emerging across tech. For instance, Jensen Huang, CEO of Nvidia, commented he’d be “deeply alarmed” if a $500,000 developer spent less than $250,000 on AI tokens. Likewise, token leaderboards found their way to big companies like Meta and Disney .

Yet despite its impressive momentum, the practice has also proved to be controversial. After all, spending tokens doesn’t provide value to the business directly. April Zheng, who leads GenAI strategy for Agentforce on Salesforce Help, told me that tokenmaxxing is both a vanity metric and genuinely impactful which makes it “hard to dismiss outright.”

“Tokens are the only broadly available signal right now for whether someone is actually using AI. But measuring productivity by token consumption is like judging how hard someone worked by whether their car was in the parking lot at 10pm. It might mean they’re heads down doing great work. It might mean they went to happy hour and got too drunk to drive home. The metric can’t tell the difference,” Zheng told me via email.

In any case, tokens do provide a metric to verify that developers are using coding agents. “It’s a great measure of input, but not a great measure of output or outcomes,” Dheeraj Pandey, cofounder and CEO of enterprise platform DevRev and former Nutanix CEO, told me in a video interview.

Pandey also shared concerns over costs, warning that companies may need to reduce token consumption significantly, particularly if they want to reduce payroll. “You can’t reduce payroll if the tokens are costing more than the payroll itself. So one has to really think as an entrepreneur to add up payroll costs and token costs and say, ‘Have we got more productive, or are we just burning and spending a lot of money on tokens?’ Because it’s a fashion fad right now," Pandey said.

High token consumption is under fire because many companies aren’t tying token consumption to tangible business impact. The lack of connection between the two naturally draws scrutiny, given the high cost of compute, and makes tokenmaxxing very difficult to justify.

“The ‘fire sale’ for AI is nearing its end. Companies no longer need to subsidise compute costs to drive adoption, and the era of ‘tokenmaxxing,’ is likely fading with it. Compute is expensive, and in many cases it’s more efficient to do the work directly than route everything through an AI model,” Cassidy Williams, senior director of developer advocacy at GitHub, told me via email.

Instead, Williams says “the conversation is moving from how to use as many tokens as possible for a fixed cost to how to use a budgeted amount of compute to deliver the most meaningful value.”

Kylan Gibbs, CEO of AI research lab Inworld AI, also noted increasing financial scrutiny in the industry. “The companies that confused token consumption with engineering will spend 2026 explaining their AI line items to a CFO who has stopped finding the dashboard charming,” Gibbs said in an email.

The fact that Microsoft is phasing out Claude Code highlights a deep rooted concern over token expenditure. “Microsoft canceling Claude Code is the first time a hyperscaler with P&L on both sides of the trade has walked away from a competitor’s tool its own engineers preferred. When the operator with the deepest visibility into the actual unit economics chooses cost discipline over developer preference, the marketing campaign is over,” Gibbs said.

As one of the largest software enterprises in the world, IBM has been building throughout the AI coding craze, but it hasn’t resorted to tokenmaxxing. “Tokenmaxxing as a measure of productivity is basically flawed,” Neel Sundaresan, GM of automation and AI at IBM and leader of the team building IBM Bob, told me in a video interview. “It is just one of the many metrics we should use to measure productivity.”

One of the key problems with tokenmaxxing, according to Sundaresan, is that it can be gamed. So instead of relying on monitoring token consumption, Sundaresan looks at a range of metrics, including the quality of the code generated, lines of code, plus the number of files touched, PRs, bugs fixed, and PR reversals.

In terms of spending, he says that the average developer at IBM doesn’t use more than $150 per month, but notes there are a number of power users spending $1000 a month or more. He also added that the company’s AI development platform Bob helps to control costs by selecting what models to use, choosing smaller models for small tasks and larger models for bigger tasks.

Instead of mandating AI use, IBM has achieved broad adoption with over 95% engagement by giving employees the power to choose whether or not to use AI. “Within IBM, we really don’t just say ‘hey, you have to use AI,’ we don’t even force our developers to use AI. Use it if you need it. If you don’t use it, probably either you do not know how to use it or AI is not useful for you,” Sundaresan said.

As the tokenmaxxing trend appears to subside, other approaches are emerging in its place. “We are seeing valuemaxxing emerge as a necessary market-wide reset on how organizations manage technology spend and risk in an AI-driven world,” Becky Trevino, chief product officer at IT management software provider Flexera, told me via email.

“Valuemaxxing challenges the need for more tools, more tokens, or more spend and pushes teams for more details on return, accountability and control,” Trevino said. It "also shifts the focus from spending more on technology to show we’re progressing on our AI transformation to making smarter and more intentional investments that drive real ROI,” Trevino said.

These kinds of approaches are a reflection that rolling out at AI isn’t necessarily a silver bullet, and providing employees with access to Claude Code isn’t going to make them more productive by default.

Trevino argues that getting value from AI comes down to setting up more stronger and more structured oversight, eliminating redundant applications, demanding visibility into tech usage and avoiding shortcuts.

Many companies have aggressively adopted AI into their workflows, but the high costs of frontier AI models highlight the need for greater governance, or costs can quickly spiral out of control.