Until recently, corporations encouraged employees to use AI without restrictions — tokens were handed out as bonuses, and internal rankings prioritized those who "burned" the most. Today, that same euphoria has turned into panic: access to models is being cut off, strict limits are being imposed, and budgets are being urgently revised.
Let's break down how tokens went from "free fuel" to the main item of unforeseen expenses and why corporations missed this moment.
What is a token and why do you pay for it
A token is the minimum unit of memory that a language model processes. Depending on the type of LLM, one token can correspond to several characters or an entire word. On average, 100 words are broken down into 130-140 tokens. They are divided into input (user request) and output (model response).
For a long time, tokens were perceived as an abstract unit of information in neural network API systems. Everything changed in March 2023 when OpenAI released GPT-4 and established public token-based pricing: $30 per 1 million input tokens and $60 per 1 million output tokens. From that moment, tokens became the unit of cost measurement for all AI tools.
From 2023 to 2025, there was an "era of subsidization": corporate developers paid for tokens, but prices were soft. However, since the spring of 2026, all major providers have fully transitioned to monetizing requests. In April, Anthropic made this decision, followed by GitHub Copilot in July. The current period is already being called the beginning of the "tokenapocalypse" era.
Price drops, demand rises: Jevons paradox in action
Over the past three years, the actual cost of tokens has decreased by approximately 98%. However, according to industry analyses, enterprise AI spending has surged by 320%. The average annual AI budget has grown from $1.2 million in 2024 to $7 million in 2026. The reason is an explosive increase in usage volumes, especially due to agentic technologies.
A simple linear workflow in 2023 cost about $0.04 per interaction. A request to a full-fledged agentic system already costs $1.2, which is 30 times more. For the most complex tasks, costs increase from 50 to 500 times. Economists call this the Jevons paradox: a decrease in the price of a resource unit leads to an increase in its overall consumption. In the case of AI, the effect turned out to be hypertrophied.
Who has already paid the price: cases of Uber, Microsoft, and Amazon
Uber. The company actively implemented AI tools Claude Code and Cursor. In May, CEO Dara Khosrowshahi boasted that autonomous agents write 10% of all code. But just a month later, Uber introduced a limit of $1500 per month per employee. The CTO admitted that he personally spent $1200 on tokens during a single two-hour demo, exhausting the annual programming budget.
Microsoft. In December 2025, the corporation granted thousands of developers access to Anthropic's Claude Code. Six months later, the experiment was shut down: most licenses were revoked, and developers were moved to Copilot CLI. According to analysts, the abandonment of Claude is a signal that the economics of corporate AI coding at current token prices do not work. The tools are so good that they are used constantly, and this is bankrupting companies.
Amazon. In June, a restructuring of the partnership with Anthropic was announced: the corporation will now pay not for compute hours, but for tokens consumed. Experts believe this could increase Amazon's costs for Claude models, although the company denies this.
Costly mistakes and the "tokenapocalypse"
Examples of excessive spending are becoming increasingly alarming. One company forgot to set limits on Claude Opus usage for employees and received a $500 million bill at the end of the month. Another developer spent over $40,000 on tokens in a single month. According to TechCrunch, token consumption per developer increased 18.6 times over nine months, mainly due to agentic features.
At the same time, the return on these costs remains questionable: most companies still cannot measure the ultimate commercial value of code produced with AI.
Not the foggiest future
The term "tokenapocalypse" clearly exaggerates the problem. Large corporations like Microsoft and Amazon can easily "pay the bills," while smaller companies can adapt if they consume wisely. Against this backdrop, the Linux Foundation presented the Tokenomics Foundation project — a new standard that will introduce "discipline" in the context of AI costs, by analogy with FinOps for cloud services.
My view: Tokens have already become a new budget item for any enterprise striving to keep pace with technology. Ignoring this fact means risking financial stability. As AI develops, such expenses will become as predictable as costs for basic corporate subscriptions, but for now, the market is going through a painful phase of transition from euphoria to common sense.