Claude credits vanish after AI guardrails slip

A company reportedly burned through roughly $500 million in Claude credits in a single month after it failed to set usage limits for employees—an incident that is feeding a wider corporate pushback over rising AI costs, “tokenmaxxing,” and the gap between prom

In the space of a month. an unspecified company reportedly burned through roughly $500 million in Claude credits after forgetting something basic: guardrails for how employees could use the system. It wasn’t a slow burn or a gradual overspend—this was the kind of mistake that turns an AI experiment into a line item that finance teams can’t ignore.

The episode is being treated as more than a cautionary tale. It lands in the middle of a broader shift in sentiment about whether AI is actually delivering the enterprise cost relief it once promised. Corporate leaders have been growing louder about the mismatch between what they hoped AI would produce and what they’re seeing in practice—especially when mounting API costs meet outputs that don’t reliably translate into productivity gains.

That pushback is showing up in comments and hiring choices across the industry. Leaders at brands such as Costco. Delta Airlines. and IBM have echoed concerns about AI use while favoring the human workforce. even as other major players—Amazon. Meta. and Microsoft—continue to cut jobs. Then came additional friction from the business side: Uber’s new COO. Andrew Macdonald. weighed in on AI-related costs and token usage that. in his framing. isn’t improving workers’ productivity as it should. The reaction online was largely positive. It was followed by reports that Uber engineers had already exhausted their AI budget for 2026.

Against that backdrop, the “tokenmaxxing” phenomenon is starting to lose support inside companies. The term refers to the tendency to burn through AI credits as fast as possible. As this mindset spreads—and then becomes expensive—executives are reportedly pushing back against the idea that more usage automatically means more value.

The economics of the problem are getting harder to explain away. AI providers. including Google and Anthropic. have moved toward usage-based billing and stricter usage limits. a shift that has sparked agitation among non-enterprise users. More broadly. a recent Gartner report says inference costs for generative AI models in 2030 will be only a tenth of what they were in 2025. But that doesn’t end the debate. because the same report also suggests token usage could grow anywhere from 5 to 30 times the current level—especially as companies rely more on AI agents and build systems that push tokens through more complex workflows.

Even companies that are betting their future on AI are adjusting how far they let it go. Earlier this month. Microsoft reportedly began canceling Claude subscriptions and discouraging employees from using it too much. just six months after it began pushing various workers across different profiles to “vibe-code.” The message from the internal changes is familiar: stop treating AI like unlimited capacity and start treating it like a budget.

The $500 million Claude-credits incident reads like the latest proof that the original promise—lower costs through automation—can collapse under real-world usage. If employees are given unbounded access, “enterprise savings” can turn into enterprise billing. None of that guarantees a reversal of the AI push. Still. the direction of travel is clear enough: more companies are expected to budget AI usage tightly. restrict it to certain activities. and place the kind of guardrails that prevent a single month from wiping out what was meant to last.

The industry’s AI fervor may not break. But the AI dream—at least the version built on free rein and effortless cost control—may already be ending.

AI costs Claude credits token usage tokenmaxxing Anthropic Google Microsoft Uber COO Andrew Macdonald Gartner inference costs generative AI