Token Speed Converter
Convert tokens per second into words per minute, pages per hour and time-to-read.
Reference — measured model speeds
Median public benchmark figures, not vendor guarantees. Click any model to load its speed into the converter.
at 80 tokens/sec — about 432 pages/hour (500 words/page). Generating 10,000 tokens takes 2m 5s.
- Words per minute
- 3,600
- Words per hour
- 216,000
- Pages per hour
- 432.0
- Time for 10,000 tokens
- 2m 5s
For scale: skilled human typing is ~80 wpm and silent reading ~250 wpm. At 80 tok/s a model writes about 14.4× faster than you can read it.
Conversion assumes English prose at ≈0.75 words/token; code and CJK text convert differently. Time-to-generate excludes time-to-first-token latency.
How it works
Vendors quote model speed in tokens per second, but nobody plans work in tokens per second. You plan in human units: how many words per minute is that, how many pages of documentation per hour, how long until this 10,000-token report finishes streaming. This converter does the unit translation both directions and anchors it with measured speeds for every model in our dataset.
The math is short and stated. English prose averages roughly 0.75 words per token (the flip side of the familiar 4-characters-per-token rule), so words per minute is tokens per second × 60 × 0.75. Pages use the 500-words-per-page manuscript convention. Time-to-generate divides your target token count by the speed — pure sustained throughput, excluding time-to-first-token, which is a separate latency budget that dominates short replies.
The reference table converts every model's benchmark speed into the same human units, and clicking a row loads that speed into the converter. The figures are median public benchmark measurements from our verified dataset, not vendor guarantees — real throughput moves with load, region and output type, so treat them as planning midpoints. The spread is wide enough to matter: the fastest small models stream at several times the rate of the largest frontier ones.
What the numbers teach: almost any model out-writes human reading speed (about 250 words per minute, a mere 5-6 tokens per second), which is why streaming chat feels instant on every tier. Speed differentiates on bulk generation — documentation runs, large-scale summarization, agent sessions producing thousands of tokens per turn — where the gap between 45 and 200 tokens per second is the gap between an overnight job and a coffee break.
Caveats, honestly: the words-per-token ratio is an English-prose average; code runs closer to three characters per token and converts differently, and CJK text differs again — the Multilingual Token Ratio tool measures that directly. For end-to-end latency including time-to-first-token, see the Streaming Latency Estimator; for what those generated tokens cost rather than how long they take, the Token Cost Calculator prices them per model.
Frequently asked questions
How do tokens per second convert to words per minute?
English prose averages about 0.75 words per token, so multiply tokens per second by 60 and then by 0.75. A model streaming at 80 tokens per second writes roughly 3,600 words per minute — more than fourteen times typical silent reading speed. The conversion is approximate: code, dense markup and non-English text all carry different words-per-token ratios.
Where do the reference model speeds come from?
They are the tokensPerSec figures from our verified model dataset — median public benchmark measurements of sustained output streaming, not vendor guarantees. Real-world speed varies with load, region, prompt length and output type, sometimes by a factor of two on the same model in the same day. Treat them as representative midpoints for planning, not SLAs.
Does time-to-generate include time-to-first-token?
No, deliberately. The converter models sustained streaming throughput only. Time-to-first-token — the pause before output begins — adds anywhere from a few hundred milliseconds to tens of seconds for long prompts or reasoning-heavy models, and it dominates perceived latency for short outputs. For a 200-token reply, TTFT often matters more than streaming speed; for a 10,000-token report, throughput wins.
Why do pages per hour use 500 words per page?
It is the standard manuscript convention — double-spaced 12-point text averages 250-500 words per page, and 500 is the common round figure for single-spaced professional documents. If your documents run denser or lighter, scale linearly: pages per hour is just words per hour divided by your words-per-page figure.
When does generation speed actually matter for choosing a model?
When output is long or a human is waiting. Interactive chat feels instant above reading speed (~250 wpm, only ~5-6 tokens/sec), so nearly every model clears that bar; the differences bite on bulk work — generating documentation, large refactors, batch summarization — where a 200 tok/s model finishes a million-token workload in under ninety minutes and a 45 tok/s model takes over six hours.
Built by FORG — AI cost observability for agentic coding. Free tools, no signup, nothing leaves your browser.
Learn about FORG