AI Token Bills Surge: 500M, 1.3M, and 18K in One Night

Burned $500 million in one month!

Recently, a massive blunder surfaced in the tech world. According to Axios, a company spent $500 million on Claude in just one month.

The reason is both absurd and frustrating: the management forgot to set a usage limit when granting employees access to Claude accounts.

In fact, AI-related bill defaults are not limited to just this one company.

In April this year, a Google Cloud user received an $18,000 bill overnight after an API key left exposed in a public service was abused, despite having only a $7 budget.

This unfortunate username, Jesse Davies, is an Australian AI consultant and founder of Agentic Labs. He set two safeguards on his Google Cloud account: a budget alert at 10 AUD (approximately $7) and a hard spending cap at $1,400.

According to Tom's Hardware, the attacker discovered a Cloud Run service the user had published months earlier, sending over 60,000 requests—both safeguards failed: billing calculations had a delay, and by the time the system responded, the charges had surged to $18,000.

In mid-May, Peter Steinberger, founder of the open-source project OpenClaw, posted a screenshot on X showing a $1.3 million OpenAI API bill over 30 days.

His team consists of only three people, but they orchestrate 100 Codex agents running in parallel: consuming 60.3 billion tokens and generating 7.6 million requests over 30 days. Fortunately, the $1.3 million in costs was not paid out of his own pocket.

Steinberger joined OpenAI in February this year, and the $1.3 million was treated as an internal experiment:

Testing without considering token costs, how far can AI programming go? He added that this result comes from Codex's "Fast Mode" (high-speed billing); turning it off would cost approximately $300,000.

Earlier, Uber’s CTO, Praveen Neppalli Naga, also admitted to The Information that the company had exhausted its entire year’s Claude Code budget by April, and their COO publicly stated that AI costs are becoming increasingly “hard to justify.”

500 million, 1.3 million, 18 thousand—although the amounts differ by several orders of magnitude, they all point to the same fact:

In the age of agents, a lost key, an army of agents working around the clock, or an account without a spending limit—any one of these can wipe out your Token balance overnight.

The answer lies mainly in the change in billing methods.

Since April this year, OpenAI's subscription pricing model has shifted to usage-based billing by tokens.

On April 2, Codex changed its billing from message-based estimation to alignment with token usage, separating charges for input, cached input, and output tokens. On April 23, this pricing model was extended to all Enterprise, Edu, Health, and Gov plans: the invisible discount within the monthly fee was removed.

GitHub has also followed suit, recently announcing that all Copilot plans will transition to usage-based billing starting June 1, 2026. The previous premium request logic will be retired and replaced with AI credits, billed according to actual consumption of input tokens, output tokens, and cached tokens, based on the API rates for each model.

GitHub officially explained the reason for this:

A quick chat question now costs the user the same as a self-directed coding task running for several hours. GitHub has been subsidizing users running heavy tasks, but this model is no longer sustainable.

Before the rise of AI agents, the costs of chat and completion were similar, making the monthly fee manageable.

After the rise of agents, a single task can run for hours and modify entire codebases, causing the cost difference between heavy and light users to span several orders of magnitude. Subscription pricing models collapse in the face of this disparity.

As soon as the news broke, there was an uproar on Reddit and X.

A developer with ID JBusu shared a screenshot of their bill, calling the new pricing "a joke." Their previous monthly cost of $28.12 would rise to $746.01 under the new rates; they’ve decided to cancel, saying, "It’s cheaper for me to rent a cloud server myself."

Another user shared a screenshot showing an even more extreme case, with fees skyrocketing from $50 to $3,000. He expressed surprise at the outrageous pricing, asking, “Does anyone still subscribe?”

However, some long-time Copilot users have come forward to counter: these extreme bills are mostly the result of vibe coders who don’t care about burning tokens, and they don’t necessarily reflect normal usage.

A long-time user commented: “I use it all day, and I rarely go over my limit at month-end—it’s hard to believe this is due to differences in work complexity.” Another user was more direct: “Some people just want an全自动 YOLO mode where AI runs wild. Removing this waste actually benefits everyone else.”

One thing must be clear: GitHub has not eliminated its monthly fee, and the base subscription price has not changed. What has truly changed is that additional usage, agent tasks, and more expensive model calls are now subject to usage-based billing.

Those most affected are heavy agent users who rely on Copilot to run long-chain tasks.

Monthly fees have slipped, as the platform changed its billing rules, while users relying on AI are also burning through funds rapidly.

In May, Business Insider reported that Amazon took down an internal AI usage leaderboard called KiroRank.

The report, citing knowledgeable sources, said the leaderboard has quietly fostered an unusual work pattern: some employees artificially inflate their rankings by consuming tokens on tasks that don’t solve real problems, solely to climb higher on the list.

After the incident came to light, Amazon Senior Vice President Dave Treadwell directly addressed all employees: “Don’t use AI just for the sake of using it. Use it to solve customer problems, address business challenges, and drive innovation.”

Although this seems absurd, it’s not surprising—when burning tokens can make the leaderboard, employees will naturally burn tokens.

Silicon Valley coined a specific term for this phenomenon: Tokenmaxxing—treating consumption as productivity.

Axios’s report also mentioned that a CTO discovered employees using AI models to check the weather and write routine emails—simple tasks that, when paired with the most expensive cutting-edge models, caused bills to surge silently.

KiroRank is not an official Amazon evaluation system, but rather an informal tool created by employees. However, it clearly reveals a classic management principle: when KPIs are set incorrectly, people will find the most clever ways to exploit the system.

Equating "how much was used" with "how well it was done"—this is the institutional root of this round of AI waste.

On the other side of token bill anxiety, someone has quietly turned it into a business.

Path one: Feed the AI with context.

Glean is Arvind’s own company. It builds enterprise AI assistants that unify knowledge scattered across a company, enabling employees’ AI tools to access context directly—eliminating the need to search endlessly. With less detouring, AI consumes fewer tokens.

This mechanism enabled Glean’s annual revenue to triple over 15 months, surpassing $300 million, with clients including Databricks, Reddit, and Samsung.

Second option: Assign tasks to the right models.

Factory AI, a startup specializing in model routing, does exactly this: automatically assigns each task to the most suitable model—simple tasks go to cost-effective models, while complex tasks go to high-end ones. Arvind also noted: getting routing right can save up to 10 times the cost.

Both paths lead to the same goal: let AI do the work, but don’t let it waste resources.

Research in academia is also laying the foundation for this shift.

https://arxiv.org/pdf/2604.22750

A 2026 April arXiv paper first systematically broke down how agent coding tasks actually burn money.

Conclusion 1: The token consumption for agent tasks can reach thousands of times that of standard code reasoning and code conversations, with input tokens being the primary driver of increased costs.

Conclusion 2: Running the same task multiple times can result in a 30-fold difference in token consumption.

Conclusion three: Higher token consumption does not necessarily lead to higher accuracy. Precision often peaks at moderate costs—spending more yields diminishing returns.

The paper also found that state-of-the-art models cannot even accurately predict how many tokens they will consume, consistently underestimating the actual cost.

You think spending more money means getting more done. In reality, the money is spent, but the work isn’t necessarily better, and the budget is still hard to estimate.

This is the first time in my memory that technical costs have begun to equal labor costs.

On May 29, Glean CEO Arvind Jain said this in an interview with CNBC reporter Deirdre Bosa.

This is also corroborated by Bryan Catanzaro, Vice President of Deep Learning at NVIDIA.

In his Axios interview, he mentioned that for his team, the cost of computing power has far exceeded employee salaries.

A similar phenomenon is emerging across multiple companies: from Glean, which builds enterprise AI, to NVIDIA, which sells AI computing power, to Uber, which uses AI—all are reevaluating the numbers.

In Arvind’s view, historically, technology accounted for only a small portion of a company’s overall costs, but now AI costs have reached parity with payroll, and many companies burn through their annual AI budgets in just one to two months.

Over the past year, AI adoption rates were revered as a key metric: the more you used, the more advanced you seemed; burning tokens meant embracing the future. Now, many companies are reflecting on this simple question: What did those burned tokens actually deliver?

The window for free monthly access has just closed.

Next, all developers face this question: how to use resources wisely to maximize the value of each token.

The true winner in the future will undoubtedly be the one who learns to account for tokens first.

Reference materials:

https://x.com/dee_bosa/status/2060791500049613306

https://www.cnbc.com/2026/05/29/-tokens-or-humans-the-new-corporate-trade-off.html

https://www.axios.com/2026/05/28/ai-spending-roi-enterprise-costs%20

https://www.businessinsider.com/amazon-ai-leaderboard-tokenmaxxing-2026-5

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.