Measure a complete household job
Try a short list lookup and a longer summary separately. A model that answers a greeting quickly may be slow when it needs household tools. Agree on an acceptable wait before comparing machines so the fastest benchmark does not replace the actual household requirement. A greeting exercises little of the real workflow. A useful trial asks for a fictional note, waits for any tool call, and ends when a reader can act on a verified answer.
Record a small comparison
Write down the model, machine, request, whether it was the first request after startup, time to first useful text and time to completion. Repeat the same job several times. Record an incorrect or incomplete answer as such regardless of speed; a fast unsupported guess is not a successful run.
Choose the limit from the household’s use
Waiting for an occasional background summary may be acceptable even when the same delay feels unusable for a grocery lookup in a shop. Decide separately for each task. When comparing machines or models, keep the source material fixed and check whether other work was running during the measurement.