The average adult reads about 238 words a minute. Once an AI model is generating faster than that, every extra token per second is text arriving faster than a person could read it even if nothing else was happening on the page. Real recorded runs, converted to the same unit as human reading speed, below.
The most-cited figure for adult reading speed, "300 words per minute," turns out to be an overestimate. The real number comes from a 2019 meta-analysis pooling 190 separate studies and roughly 18,000 participants: adults read English non-fiction silently at an average of 238 words per minute, with most people somewhere between 175 and 300 WPM. Reading aloud is slower still, averaging 183 WPM, bottlenecked by actually saying the words.
Source: Brysbaert, M. (2019). "How many words do we read per minute? A review and meta-analysis of reading rate." Journal of Memory and Language, 109. This is a different measurement from this site's own 15-second typing test, which measures typing speed, not reading speed, they are not interchangeable.
A token is roughly three-quarters of a word in English (OpenAI's own published rule of thumb: "100 tokens ≈ 75 words," the same ratio this site's glossary uses). So: tokens per second × 0.75 × 60 = words per minute. Working that backward, an AI model only needs to sustain about 5.3 tokens per second to match the average adult's own reading speed. Below that, a reader can plausibly keep up with the words as they stream in. Above it, the model is writing faster than any reader can read, regardless of how the words arrive.
Four real measurements, picked for having a long, stable generation window rather than for being the most dramatic numbers available:
Every one of these, including the slowest, already generates faster than the average adult reads. The gap doesn't stay close either: the fastest of the four here writes at roughly 31× reading speed.
Past the reading-speed breakeven, more raw tokens per second stops changing what a person actually experiences reading the answer, it's already arriving faster than they can take it in. What still matters a great deal is time to first token: how long someone waits before anything appears at all. A model that starts instantly at a modest generation speed can feel faster to use than one that sits silent for several seconds before writing at 10x the rate. Every race on this site shows both numbers side by side for exactly this reason, speed alone is an incomplete answer. See how we measure → for the full breakdown of what's timed and how.
Every AI number above is computed from a real recorded run on this site, the same functions every other page uses. Open dataset → · Compare AI models → · Glossary →