A week of heavy Claude and Codex use reached 19.6B token operations
Claude Code and Codex were used heavily for one week across five computers, two Anthropic accounts, and ChatGPT Pro. An audit of more than 247 sessions found 19.59 billion and 61.8 million generated . At API list prices, that activity would have cost about $19,700, but the cost $600 per month and another $40 of was used for company work.
The estimated actual cost for the week was about $178, making the API-price equivalent roughly 111 times larger. About 94% of all operations were cache reads, so most of the 19.59 billion operations did not represent newly generated tokens. Saturday alone accounted for 13.83 billion operations and about $13,527 at API list prices.
The agents handled coding, audits, testing, research, , genealogy, workflow design, and fictional-language development. The largest Claude session ran for more than 22 hours and added nearly 30,000 lines; a request for five subagents unexpectedly launched 135, yet the work remained mostly coherent with little drift or architectural conflict.
Key points
- The audit covered more than 247 sessions across five computers.
- The tools produced 61.8 million , while cache reads made up about 94% of all operations.
- The API list-price equivalent was about $19,700, compared with an estimated actual weekly cost of $178.
- The longest Claude session ran for over 22 hours and added nearly 30,000 lines.
- A request for five subagents launched 135 instead.