What is a reasoning model?
Route every step through one and your agent gets slow and expensive for no reason. Skip it everywhere and the agent gets the hard steps wrong. The skill is telling which step is which.
Reasoning model
extended thinkingchain-of-thought model
A reasoning model is a large language model trained to work through steps before it writes a final answer. It does not answer straight from the prompt. That extra step trades speed and token cost for better accuracy on hard problems: multi-step math, code debugging, planning. Not every question needs it.
A standard model reads your prompt and starts writing the answer in the same pass. A reasoning model works through the problem first. It breaks the problem into parts, tries a step, checks it, and backs up if the step fails. That working-out is often hidden from the final response. It still costs real compute and real time.
Anthropic's Claude models and OpenAI's o-series and GPT-5 both support this mode. It is usually a setting you turn on per request, not a separate product. A low reasoning effort stays close to a standard model's speed. A high one can take many times longer on the same request. You trade that time for a harder problem answered more reliably.
The trade nobody states out loud
Every reasoning token is a token you pay for and a token the user waits on. Turn extended thinking on for an entire pipeline, and a request that answered in two seconds can take twenty. Multiply that by every call in a busy agent. The wait changes shape, and so does the bill.
The buyers we talk to usually discover this after shipping, not before. Picture a support agent asking "what is your order number". Reasoning through that step is slower for no benefit, because the next move is obvious. Now picture the same agent deciding whether a refund breaks a contract clause. That is where the extra steps earn their cost, because a wrong answer there is expensive.
So the production pattern skips the on-off switch. First classify the step. Then send only the steps that need depth through a reasoning model, and let a faster model handle the rest. Get that classifier wrong and you either overpay everywhere or lose accuracy where it counts.
- RequestA step in an agent or workflow.
- ClassifySimple lookup, or dependent steps?
- RouteFast model, or reasoning model.
- ReasonOnly on the hard branch.
- AnswerChecked before it ships.
Most steps in a working agent never reach the fourth node. That is by design, not a shortfall.
Related questions
01How is a reasoning model different from a regular LLM?
A regular LLM generates its answer directly from the prompt. A reasoning model is trained to produce an internal sequence of intermediate steps first, then answer from that working. Both are large language models; reasoning is a mode or a training difference, not a separate category of technology.
02When should you use a reasoning model?
Use one on steps with several dependent parts: debugging, planning, comparing conflicting sources, or any decision where a wrong answer is expensive. Skip it on lookups, formatting and simple classification, where a fast model already gets the right answer without the extra tokens.
03Does a reasoning model cost more to run?
Yes. The intermediate steps consume tokens even when they stay hidden from the final response. A higher reasoning effort setting can take many times longer than a standard call. That cost is why production systems route only specific steps through one, instead of using it everywhere.
04Can you see a reasoning model's intermediate steps?
Providers differ on this. Some show a summary of the reasoning; others keep it internal and return only the final answer. Either way, the reasoning tokens get billed, whether or not you can read them.
