Usage & Prompting
Token
Models don't read text, they read tokens: fragments of roughly three to four characters. “Incomprehensible” is several tokens, “and” is one. Everything is measured in tokens: the length of the context window, speed, above all price. Anyone who wants to understand AI costs has to think in tokens, not pages.
An example
Models don't read words but fragments. “Company” is usually one token, “Grundstücksverkehrsgenehmigung” is several, and a number like 2026 can split into two. German systematically needs more tokens than English for this: for the same content often thirty to fifty percent more. Anyone who calculates costs from English sample calculations and then processes German texts will be off accordingly.
A closer look
Text is broken into pieces that occur more often than single letters and are shorter than words. In German that is roughly 0.7 words per token; technical terms, umlauts, and code need more. Prices and context limits are counted in this unit.
Why it matters
Tokens are the unit of account for almost everything practical: price, speed, and the question of how much context fits. A reliable calculation therefore needs no estimate but a measurement on real documents of your own; providers offer counters for this. The second effect is subtler: because numbers and rare technical terms are split awkwardly, models are weaker on exactly these. Copying a material number wrongly is not a lapse of concentration but a consequence of the splitting.
Commonly misunderstood
Because the model sees tokens and not letters, tasks such as counting letters or writing words backward are unexpectedly error-prone.