Just a year ago, integrating artificial intelligence into the corporate environment resembled a race without rules: companies handed out access to models like corporate bonuses, encouraging employees to "burn" tokens. Today, the same picture looks different — executives are panicking, revoking access, imposing strict limits, and hastily revising budgets.

The token, which was once just an abstract unit of data processing, has turned into the main line item for unforeseen expenses. Let's figure out why this happened and how corporations were caught off guard.

From "Free Fuel" to "Token Apocalypse"

In the context of large language models (LLMs), a token is the minimum unit of information processed by the neural network. One token can correspond to part of a word or a whole word, and 100 words are roughly 130–140 tokens. They are divided into input tokens (user prompts) and output tokens (generated responses).

For a long time, tokens were perceived as a technical detail with no financial significance. Everything changed in March 2023 when OpenAI released GPT-4 and set public pricing: $30 per 1 million input tokens and $60 per 1 million output tokens. From that moment, gradual monetization began, culminating by spring 2026 in a complete transition to pay-per-use.

Analysts have already dubbed the current period the "token apocalypse" — companies are forced to urgently include additional costs in their budgets that were previously simply ignored.

Jevons Paradox in Action

The most interesting aspect of this situation is the paradox: over three years, the average cost of an input token has dropped by about 98%. However, corporate spending on AI has skyrocketed by 320% — from $1.2 million in 2024 to $7 million in 2026. The reason is an explosive growth in usage volumes, especially driven by agentic technologies.

A simple query in 2023 cost about $0.04. A query to a full-fledged agentic system already costs about $1.2, which is 30 times more. For complex workflows, the gap reaches 500 times. We are witnessing the classic Jevons paradox: a decrease in the price of a resource leads to a hypertrophied increase in its consumption.

Case Studies: Who Has Already Been Burned

Uber. The company actively implemented Claude Code and Cursor, with CEO Dara Khosrowshahi boasting that agents write 10% of the code. However, it soon turned out that the annual programming budget was exhausted in just a few months. One CTO spent $1,200 on tokens during a two-hour demo. Now Uber has introduced a limit of $1,500 per month per employee.

Microsoft. In December 2025, the corporation gave thousands of developers access to Anthropic's Claude Code. The experiment was so successful that engineers preferred it over their own GitHub Copilot. Six months later, Microsoft shut down the program, citing the need to migrate to Copilot CLI. Journalists suggest the real reason was exorbitant token costs.

Amazon. The company is restructuring its partnership with Anthropic, switching from paying for compute hours to paying for tokens. Experts believe this could increase Amazon's costs for Claude models, although the corporation itself denies this.

Especially telling is the story of an anonymous company that forgot to set limits on Claude Opus usage and received a $500 million bill. Additionally, one developer spent over $40,000 on tokens in a single month.

Conclusions and Forecast

Large corporations will undoubtedly survive the "token apocalypse" — Microsoft and Amazon have direct investments in Anthropic and OpenAI. But for more modest firms, this could become a serious financial blow.

Against this backdrop, the Linux Foundation has announced the Tokenomics Foundation project — a new standard designed to introduce "discipline" into AI cost management, analogous to cloud FinOps. My professional forecast: tokens will become as routine a budget item as cloud service subscriptions. But until then, companies will have to learn to count every token — or risk getting a bill that negates the entire economic efficiency of implementation.