tokenspeed

How fast is 30 tokens per second, really?

Every local-LLM benchmark reports throughput: "47 tok/s on an M3," "180 tok/s on a 4090," "500 tok/s on Groq." But unless you've actually watched tokens stream at those speeds, the numbers are hard to internalize.

tokenspeed is a tiny terminal toy that streams fake tokens at any rate you set, so you can see what those numbers actually look like.

Three modes:

• code — syntax-highlighted pseudo-code (Python/Rust/JS), the most common thing you watch stream out of an LLM
• text — Wikipedia prose, for the chat/answer case
• think — dim-italic reasoning sentences alternating with code, mimicking a reasoning model thinking out loud

If you don't pick a mode on the command line, it'll ask.

Run it

python3 tokenspeed.py                  # 30 tok/s, prompts for mode
python3 tokenspeed.py 60               # 60 tok/s, prompts for mode
python3 tokenspeed.py --mode code      # skip the prompt
python3 tokenspeed.py 120 --mode think # both at once

No dependencies — just Python 3 and a real terminal.

What to try

Start at the default 30 and read along. Then hit 1 (5 tok/s) — Raspberry-Pi-class local model. Then 5 (60 tok/s) — typical hosted Claude or GPT. Then 7 (200 tok/s) — Groq territory. Then 9 (800 tok/s) — Cerebras-class, where the bottleneck is your eyeballs.

Now switch between --mode code and --mode text at the same rate. The difference is striking and intentional — see below.

Controls

Key	Action
+ / -	Nudge the rate by ×1.25
1–9	Jump to a preset: 5, 10, 20, 30, 60, 100, 200, 400, 800 tok/s
space	Pause / resume
q	Quit

What counts as a token

tokenspeed approximates BPE-style tokenization. It is not trying to exactly reproduce tiktoken, Claude's tokenizer, or any vendor-specific encoder — those disagree with each other in the details anyway.

Roughly: short words are often one token; longer identifiers frequently split into multiple chunks (e.g., processUserInput → process + User + Input, calculate_score → calculate + _score); and punctuation and operators usually count as tokens too.

The point worth internalizing: code is more token-dense than prose, so the same tok/s can feel very different depending on what's streaming. 30 tok/s of code lands far less visible content per second than 30 tok/s of English. The benchmark number is honest; the perceptual effect just varies a lot by content type, which is exactly the gap this tool exists to expose.

(English prose averages ~1.3 tokens per word, so 30 tok/s ≈ 23 words/s.)