Model cards, GPU spec sheets and serving benchmarks all report tokens per second, and almost nobody knows what any of those numbers mean in the body. Forty is fine. Four is unusable. Four hundred is indistinguishable from four thousand. Move the slider below and the text streams at exactly the rate you picked — the same accumulator a real streaming client uses, no fake typewriter easing.
This simulator needs JavaScript. The text it streams says: a token is a fragment, not a word — which is why generation rates and reading speeds do not convert one-to-one.
The highlighted fragment is the token currently being emitted. Words break mid-word on purpose — that is what a tokenizer does, and it is why "tokens per second" is not "words per second".
The red marks on the slider are the two rates we measured on our own machines on 2026-08-19. The grey ones are conversions and published third-party figures, and are labelled as such. We do not have an inference ASIC to benchmark, so we do not pretend the number is ours.
On the medians, the hosted model wins: 57 against 40. That is the number a vendor puts on a slide. But our slowest local run was 39.5 tokens a second and our slowest cloud run was 6.6 — below speech, well below reading, the rate at which a person gives up and refreshes. A pipeline that runs unattended overnight is sized by its worst run, not its median, and the machine with no worst case is the one you can schedule around.
That is why the pipelines behind this site generate locally first and reach for a hosted model only where it demonstrably pays. Not ideology — an eleven-times tail that hits at 3 a.m. with nobody watching.
→ What the same box costs to run → Get inference sized for your worst case → How the rest of it runs
The AI news that matters — in your inbox by 07:30 CET. Free, no spam.