five hundred million dollars
Output tokens cost much more than input. Most real workloads read a lot and write a little.
For that budget you could process
— tokens
— input · — output
That's like feeding Claude
— Linux kernels
Split it across a team
Spread the same budget across a team of developers over a number of months to see what each person would chew through per month.
How long the budget is spread over.
Where this came from
This started as a lark. In May 2026, Tom's Hardware reported that a mystery company accidentally spent about $500 million on Claude in a single month after failing to set usage limits on its employee licenses. That number is almost impossible to picture — so this page tries to turn it into something you can actually feel: how many tokens is that, and what would it look like spread across a team?
How the comparison sizes are estimated
These are deliberately rough, order-of-magnitude figures meant to be relatable, not precise token counts:
- Codebases — line counts from recent checkouts, at roughly 10 tokens per line: the Linux kernel (~37 million lines) ≈ 400 million tokens, Kubernetes (~3.7 million lines) ≈ 37 million tokens, and the Docker engine (moby, ~380k lines) ≈ 3.8 million tokens. Codebase counts exclude vendored third-party dependencies.
- Books and Wikipedia — taken from published word counts and converted at about 1.3 tokens per English word (Wikipedia ≈ 6 billion tokens; the Bible, Shakespeare, and the Harry Potter series are each roughly 1 million tokens).
Token-to-text ratios vary by tokenizer and content, so treat every comparison as a ballpark.
This site was built with Claude Code, so just assume everything is wrong. It took approximately 13.2 million tokens.