How the benchmark numbers are measured
Every number on the benchmarks page comes from one source and one formula. This page states both, plus the limits, so a figure can be checked or corrected rather than taken on trust.
Where the data comes from
Tokens per second and prices are taken from ModelDock benchmarks. Licence terms: Factual benchmark medians and dated street prices; reproduced with attribution. ModelDock benchmarks
How each figure is calculated
- Daily output = reading speed × hours you run the machine.
- Power cost = watts × hours × electricity price.
- Cost per day = power cost + hardware price spread over its useful life.
- Payback days = rig price ÷ money saved per day versus an API.
Assumptions behind the table
The table uses these defaults. Change any of them on the benchmarks page and every derived figure updates.
| Share of output that is reading | 90% |
|---|---|
| Hours per day | 24 h |
| Electricity price | 0.15/kWh |
| Hardware life | 5 years |
| Host allowance | 1200 |
| API reference price (read / write) | 0.1 / 0.64 $ per million tokens |
What these numbers do not cover
- Consumer GPUs only. Server parts and multi-GPU configurations are excluded.
- One workload: a sequential single-stream mix. Batched, concurrent serving runs faster in aggregate and slower per stream.
- Prices are dated street prices and move independently of the measurements.
- A payback figure is an estimate, not a quote: it assumes the machine is busy at the stated hours.
How current this is
The table was last synchronised on 2026-10-07. Measures are re-published in place, so a figure quoted from an older copy can be checked against its date.
Citing these figures
The dataset is free to cite with attribution to Strata and the underlying source. Link to this page when citing a method, and to the benchmarks page when citing a number.