Reasoning model

When you will meet it

You need this the moment a model menu shows a "standard" and a "thinking" variant with visibly different price and speed. Without it you make two symmetric errors: running simple tasks on a reasoning model (bigger bill, longer latency, same answer), or forcing a plain model through multi-step reasoning (errors snowball across steps, and you blame the model).

Minimal example

同一個多步問題，兩種模型的帳單結構（示意）：

普通模型：
  輸入一段 prompt → 輸出一段答案 → 結束

思考型模型：
  輸入同一段 prompt → 先輸出一大段思考 → 再輸出答案
  （思考內容也是輸出 token，照樣計費；
    多數平台不讓你看思考全文，也不保證留到下一輪）

→ 同一次請求，計費的輸出量可以差上一大截。
→ 換來的價值只在「多步、易錯、答錯代價高」的任務上兌現。

Note the three facts in parentheses: thinking is billed as output, usually hidden from you, and not guaranteed to persist across turns. You are paying for "a more likely correct answer", not for a reusable reasoning artifact.

What people get wrong

• Assuming reasoning models are "strictly better" and routing all traffic to them. On simple extraction, classification, small talk or latency-sensitive paths, the extra thinking rarely improves answers while genuinely doubling cost and latency. The right pattern is tiering by task: hard goes to reasoning, simple goes to standard.
• Conflating "reasoning" with "inference" — a chronic problem in Chinese sources where both get called 推理. Inference is running any model to produce output; reasoning is a training style where the model thinks before answering. Disambiguate from context; this site renders the latter as 思考型 (thinking-style).