A million tokens is fewer words than you think: we measured it

[Redaktion]

One million tokens sounds like enough space for everything. We calculated how many German words that actually is, how much a single query costs for the entire GPT-5.6 Sol collection, and when Obsidian or Notion are still worth it.

https://www.notebookcheck.net/A-million-tokens-is-fewer-words-than-you-think-we-measured-it.1352830.0.html

[Correcto mundo]

The widely cited rule of thumb is that one million tokens equals 750,000 words.

Correcto mundo:

~32000 tokens (= ~24000 words)

(1 token = 0.75 words)

(and I think way before that somewhere in other topic's comments)

Not sure why you'd create a whole new topic about this, but ok. Maybe instead rewrite
notebookcheck.net/Your-own-ChatGPT-offline-AI-without-the-cloud-on-your-laptop.1341872.0.html
based on comments, because recommending Qwen3-8B (only an outdated AI/LLM would recommend Qwen3.0, typically LLMs are behind when it comes to news). I don't want to say recommending any 8B is just a no-go, as an 8B model can be used for specialized tasks, but at the very least Qwen3.5-9B exists, but even then I'm not sure I'd recommend it. Say that one should try to get at least a 3- to 4-bit Qwen3.6-27B quant running, and/or say that one should not expect too much from any 8-9B model [as long as there have not been any major breakthroughs in LLM architectures].
PS: Qwen3.8-27B, and other sizes, will be released next week.

[Jet]

Excellent write up and clear explanation of the AI inner workings and billing systems.