$ claude-meter --month 2026-08 --source ~/.claude/transcripts
What one month of AI-assisted coding actually costs
Measuring the “local models are free” claim against 14.69 billion tokens of one person’s own usage: 71,350 requests in the 30 days ending August 28, 2026, which would have cost $14,914 at API list rates but ran on a $200-a-month subscription. The same work is priced against six home machines — two used RTX 3090s, an RTX 5090 workstation, a Strix Halo and a DGX Spark at 128 GB, and Mac Studios at 256 and 512 GB — with hardware, capital, power and 13.7 hours a month of upkeep counted.
- 14.69B
- tokens metered, 96% of them cache re-reads
- 71,350
- requests
- $200 / $14,914
- paid on the plan / metered at list price — 74.6× leverage
- 3 of 6
- desk machines cannot finish the month’s work inside a month; 4 of 6 once rejected work is redone
- 27B-class
- the only model class that keeps pace — and it is below the task
- 13.7 h/mo
- of setup and upkeep: the real bill, at whatever your hour is worth
SECTIONS
- 01 The month — where 14.69 billion tokens actually went
- 02 Machines — which desk setups could physically keep up, retries included
- 03 Bill — cost per million tokens with your hours priced at a wage you choose
- 04 Quality — benchmark scores for what the plan serves and what a box can hold
- 05 Fair fight — the same month re-billed at hosted open-model rates
- 06 Five years — cumulative cash, with breakeven months
- 07 Prints — six plates after the data portraits of W. E. B. Du Bois, Paris Exposition, 1900
SOURCES
Models, source data and every poster iteration on GitHub · the long-form companion report · the A3 nomograph poster
measured july 29 – august 28, 2026 · open-model rates checked aug 31, 2026 · © studiolo.svetim.fm