An OpenAI engineer burned through 210 billion tokens in a single week, which is 33 times the entire content of Wikipedia. At Anthropic, one Claude Code user racked up a bill of 150,000 dollars in a single month. For comparison, a student writing an essay with AI uses around 10,000 tokens.
What is tokenmaxxing?
The term describes the race to consume as many AI tokens as possible inside tech companies. The principle: use the most tokens you can, as fast as you can, and be seen doing it.
The practice became possible thanks to agentic coding tools such as Claude Code and OpenClaw. These systems run autonomously for hours, with agents able to spawn sub-agents of their own. Some run around the clock, consuming tokens continuously.
The corporate status game
Meta, Shopify and OpenAI require intensive AI use, and assess it in performance reviews. Managers reward the heavy consumers and warn those who fail to adopt the technology.
Token budgets have become a benefit in kind, on a par with a meal allowance. Internal leaderboards rank engineers by consumption. Tokens are no longer an infrastructure metric; they have become a measure of professional performance.
The problem: quantity is not quality
Tokenmaxxing reproduces presenteeism: staying late at the office then, burning tokens now. An autonomous agent can generate 10,000 lines of code of which only 200 are genuinely usable.
Using AI well
The real AI skill is precision, not volume. A well-built prompt gets the right result in 500 tokens; a vague one burns 50,000 for a mediocre result.
AS3P teaches the C.A.R.T.E.L framework:
- Context
- Audience
- Role
- Task
- Example
- Layout
The professionals who will succeed are not those consuming the most tokens, but those getting the best results from the fewest. That is the difference between using AI and mastering AI.
Sources: New York Times (20 March 2026), Economic Times HR.