i finally priced the work running on my mini
the paid meter was running at $42.50 a day. i moved routine work onto a standard M4 mac mini with 24GB, and paid frontier burn fell to $3.90 a day. then i wanted to know what the box had actually been chewing through at api prices.
1. count the calls before pricing
over 16 days the local model took 42,124 calls. that is about 2,633 a day.
2. keep prompt and output separate
the prompt side was 44,049,629 uncached tokens. the model generated 7,751,250. folding those into one number hides the part of the bill that grows fastest.
3. put list price on the load
that 16-day load at Sonnet 5.5 API list price comes to $165.61. as a run rate that is $10.35 a day, or $310.52 per 30 days.
4. watch what the cache changes
a 2,048-token prefix stays pinned in KV cache with an 88.4% hit rate and 3.5s incremental re-bakes. that helps the local box run. the api comparison above still uses the measured uncached prompt count, so the cache gets no credit.
5. leave room for paid escalations
the local model still hands off work it cannot carry: 64 escalations to paid or bigger models in 12 days, 90,623 tokens total. the $3.90 a day paid burn is measured separately from the list-price math.
6. remember what ran the tokens
a 35B-total mixture-of-experts with about 3B active parameters. on nov 2, from the 8:28 am server start to midday, it handled 471,661 tokens (395,932 prompt, 75,729 generated). none of them were billed by a cloud api.
- count the calls
- split prompt from output
- price the measured load
- check the daily run rate
- account for cache hits
- count paid escalations