Reasoning Models — Junior¶

At junior level, focus on this question:

Can you explain what a reasoning model does differently, why you pay for tokens you never see, and which tasks actually benefit?

What "reasoning" actually is¶

A reasoning model (e.g., a Claude extended-thinking class, an OpenAI o-class, Gemini thinking variants, GLM's thinking modes) generates thinking tokens before the final answer: it works through the problem — trying steps, catching its own mistakes, backtracking — then writes the answer.
It's still the same next-token loop from How LLMs Work — the model just runs many private iterations of it before producing tokens you see. Not a different kind of intelligence; more compute spent before answering.

flowchart LR P["Your prompt"] --> T["Thinking tokens<br/>hidden - but billed"] --> A["Final answer<br/>visible"]

Thinking tokens are billed as output tokens — typically several times the price of input tokens — and they add seconds to minutes of latency.
They often outnumber the visible answer by a large multiple: a 200-word answer can carry 2,000 tokens of thought.
Check the usage fields in every response: reasoning_tokens (or equivalent) is the line item that surprises people on the invoice.

Multi-step logic: math, puzzle-like problems, scheduling with constraints — tasks where the first plausible answer is usually wrong and checking steps pays.
Planning: breaking a fuzzy goal into ordered sub-tasks before acting.
Hard debugging/troubleshooting: reasoning over evidence to form and discard hypotheses.
The common thread: the task rewards thinking before answering, and the cost of a wrong answer exceeds the cost of the extra tokens.

Extraction, reformatting, classification, simple lookups, chat replies — tasks whose difficulty is in knowing, not figuring out. A standard model at temperature 0 matches quality at a fraction of cost and latency.

Reaching for a reasoning model because the task is "important." Importance ≠ difficulty; reasoning pays off only when the task needs deliberation.
Not looking at reasoning-token counts. The bill quietly runs 5–10× the visible output; nobody notices until the invoice.
Using it for latency-sensitive interactive paths. Extra seconds of thinking are visible to users waiting on a chat reply.
Assuming reasoning fixes factual gaps. If the model doesn't know the fact, thinking longer produces confidently reasoned nonsense — the gap needs data, not deliberation.

Take one task you're considering a reasoning model for; state whether it fails from lack of * figuring-out or lack of information* — the answer decides the model class.
Run one real task on both a standard and a reasoning model; compare answer quality, latency, and (from usage fields) cost including hidden tokens.
Find the reasoning-token field in your provider's response format and locate it in your most recent bill's math.

You can explain thinking tokens as more loop iterations — not a new capability category.
Any reasoning-model choice cites "needs deliberation," with a cost/latency number attached.
Hidden-token cost is computed from usage fields, not guessed.