Six commonly confused LLM API terms: context window, temperature, reasoning tokens

ClickOk5811 · reddit · 2026-10-24

A Reddit glossary untangles terms frequently confused in API questions:

Context window vs max output tokens: the window covers input + output combined; max output is a separate, smaller cap, so a huge window doesn't mean huge answers.
Tokens vs words: English averages slightly under one word per token; other languages, code and numbers cost more — tokenize before estimating cost.
Temperature vs topp: both control sampling randomness; docs recommend changing only one at a time for attribution.
Temperature 0 isn't deterministic: outputs are more repeatable but not identical; validate instead of assuming stability.
Reasoning tokens: invisible thinking tokens count toward output limits and billing and can truncate responses; some reasoning models reject sampling params.
System/developer vs user messages: different instruction-hierarchy priority, but neither is a security boundary against untrusted text.

More from coding & agent