Claude and AI agents burn through tokens faster than forecast, and businesses are learning "tokenomics"

Companies that deployed AI agents to write code have hit an unexpected problem: actual token costs came in well above forecasts, and executives are now scrambling to learn a new discipline called "tokenomics".

What happened to the budgets

Wired spoke with several companies about how they manage AI agent spending in real-world conditions. One of them, an executive at a Silicon Valley software company, described the situation with the word "insanity": his team started using Anthropic's Claude heavily for coding, and token consumption soared to levels nobody had budgeted for.

An e-commerce company is in a similar position. Its AI agents, running in the background, generate thousands of tokens on tasks that seem routine: code review, writing tests, refactoring. If a developer spends an hour on such a task, the agent "thinks" for a few minutes, but it generates tokens continuously, and the monthly bill turns out to be something else entirely.

Why agents burn tokens so fast

A token is the unit of billing with AI providers. Every request to a model and every response is charged by token count. For an ordinary chat, this is barely noticeable. But AI coding agents work in a fundamentally different way:

• before each action, they read the entire repository context
• within a single task, they call the model many times
• they generate long chains of internal reasoning before answering
• they write, test and rewrite code until they get the desired result

As a result, a task a developer finishes in an hour can "cost" tens of thousands of tokens. At the high rates charged for powerful models like Claude, that adds up to hundreds of dollars per employee for a single working day.

How companies are changing their approach

Executives are starting to introduce the concept of "tokenomics": managing token consumption the way server resources or cloud spending were managed before. The first practices have already taken shape:

• limiting the agents' context window, so the agent sees only the relevant part of the codebase
• caching repeated prompts so tokens are not recalculated from scratch
• task routing: cheap models for routine work, powerful ones for complex requests
• monitoring and alerts for abnormal spending
• revisiting the ROI of AI tools using actual costs rather than forecast ones

"We were used to thinking of AI as a SaaS subscription with a fixed price.

Now we understand it is more like cloud computing: the price depends on how much you use."

Anthropic and other providers offer tools for monitoring spending, but managing tokens remains a headache on the customer side for now.

What this means

Businesses' bet on AI coding as a way to cut development costs may not pay off if the real cost of tokens is not taken into account. Companies that master "tokenomics" first, learning to optimize consumption without sacrificing results, will gain a tangible advantage over those managing AI costs blindly.