Skip to content
AI Playbooks · 10 min read

Reasoning Models for Beginners: The Playbook on When AI Should Think Longer

Reasoning models for beginners: learn what the effort setting does, when extra AI thinking helps, what it costs, and how to pick the right level.

Reasoning models for beginners cover: an orange thinking orb above low, medium and high effort tiles

Meera is a developer in Pune. Her team’s support bot answers about two thousand customer questions a day. Last week a teammate flipped one setting, “reasoning effort”, from low to high, hoping for smarter answers. The next morning the bot felt slow, and the bill had jumped.

Meera had heard the words “reasoning model” and “thinking” many times but never really understood them. So she sat down with one question: when should an AI think longer, and when is that just wasted money? This chapter is her answer, written as reasoning models for beginners: what they are, what the effort dial does, and how to set it without guessing.

Last verified: 15 November 2026. AI tools change fast; check the linked sources for anything newer.

Who this is for: anyone who calls an AI model from code, or uses an AI app with a “thinking” or “effort” option, and wants to choose it calmly. No maths needed. Time needed: about 30-40 minutes, plus an hour to run your own small test.

Reasoning models for beginners: low, medium and high effort levels

Reasoning models for beginners: what they are, in simple words

Think of two students taking a test. The first writes the answer straight away. The second uses scratch paper first: plans, tries a step, checks it, then writes the answer. The second one is slower, uses more paper, and gets far more of the hard questions right. On easy questions, both score the same.

A reasoning model is the second student. Before it answers, it writes hidden working notes. OpenAI calls these notes reasoning tokens; Anthropic calls the same idea thinking. A token is a small piece of text, roughly part of a word (see how LLMs work).

The precise version: a reasoning model spends extra output tokens planning and checking before its final reply, and you pay for those tokens even though you usually do not see them.

What the effort setting does

The reasoning effort setting is the size of the scratch pad. Low effort means a small pad: quick, cheap, fine for simple jobs. High effort means a big pad: slower and more expensive, better for hard problems. Most providers now offer levels such as low, medium and high, and some add extra levels at either end.

One detail surprises beginners. Anthropic’s docs say effort is “a behavioral signal, not a strict token budget”. The model still thinks on a hard problem at low effort, just less. So effort is a nudge, not a hard cap.

What’s new this month

Effort is now the main control on many new models. Anthropic’s effort guide (read 15 November 2026) lists five levels, from low to max, and notes that some newer models default to medium instead of high. OpenAI’s reasoning guide lists values from none to max, depending on the model, and recommends starting at medium. Both say the same thing in different words, and it is the main lesson of reasoning models for beginners: start in the middle and measure before you go higher.

The playbook at a glance

  1. Play 1: Sort your tasks into easy, medium and hard.
  2. Play 2: Fix the prompt before you touch the dial.
  3. Play 3: Start in the middle and test low and high.
  4. Play 4: Leave room for the thinking.
  5. Play 5: Route each task to the effort it needs.

Play 1: Sort your tasks into easy, medium and hard

Goal: know which jobs actually need thinking. A simple test from beginner courses: would a careful person need scratch paper for this?

  1. List the jobs your AI feature does, for example: sort a ticket, answer a policy question, plan a refund.
  2. Mark each one. Easy: lookups, labels, reformatting. Medium: a few connected steps. Hard: many rules at once, or a costly mistake.
  3. Count how many requests each job gets per day. The big, easy jobs are where low effort saves the most.

Meera found that 80% of her bot’s traffic (illustrative) was “where is my order” style questions. Easy. Only refund disputes, mixing three policies at once, were truly hard.

You are done when: every job has an easy, medium or hard label and a rough daily volume.

Play 2: Fix the prompt before you touch the dial

Goal: do not pay for thinking to cover a vague question. More effort cannot guess what you meant.

  1. Write the goal in one sentence.
  2. List every rule the answer must follow and the facts it needs.
  3. Say what a good answer looks like, and how you would check it.
  4. Skip “think step by step”. Reasoning models already plan; a clear goal helps them more.

OpenAI’s guide says it directly: clear goals, constraints and output formats matter more than the effort setting. Our prompt engineering playbook shows how to write them.

Meera rewrote the refund prompt to list the three policies and the order they apply in. About half the “hard” mistakes vanished at the same effort level (illustrative).

You are done when: each hard task has a prompt with a goal, rules, facts and a success check.

Play 3: Start in the middle and test low and high

Goal: pick the effort level with evidence, not a hunch. Both Anthropic and OpenAI recommend running the same test at several levels. This is called an effort sweep.

  1. Build a small test set: 20-30 real questions with good answers, as in our evals playbook.
  2. Run it at medium. Write down the score, the average wait, and the tokens used.
  3. Run it again at low and at high.
  4. Pick the lowest level whose score you are happy with.

Here is what the setting looks like in code. This is an illustration in Python with Anthropic’s SDK; OpenAI’s version uses reasoning={"effort": "medium"} instead, and names change between versions.

import anthropic

client = anthropic.Anthropic()
reply = client.messages.create(
    model="your-model-name",
    max_tokens=16000,
    messages=[{"role": "user", "content": question}],
    output_config={"effort": "medium"},   # low, medium, high...
)
print(reply.content[-1].text)

In plain steps: connect to the provider, send the question, set a ceiling on total output, choose an effort level, and print the final answer.

Meera’s sweep (illustrative numbers): easy questions scored 96% at low and 97% at high, so low won. Refund disputes jumped from 70% at low to 92% at medium, and only 93% at high.

You are done when: you have a small table of score, wait and tokens for each level you tried.

Play 4: Leave room for the thinking

Goal: stop answers from being cut off. The hidden notes count toward the output limit, often called max_tokens.

  1. Remember the bill: OpenAI says reasoning tokens are billed as output tokens and take up space in the context window, the model’s working memory.
  2. Raise the output limit when you raise effort. OpenAI suggests reserving at least 25,000 tokens for reasoning and the answer to start, then adjusting.
  3. Watch for cut-offs. If the limit is hit while the model is still thinking, you can pay for tokens and get no usable answer.
  4. Log the reasoning tokens per request, so you see the cost (our observability playbook shows how).

Meera’s teammate had set the limit to 1,000 tokens. At high effort, some refund replies came back empty. That was the “bot is broken” ticket.

You are done when: your limit leaves room for thinking and you can see reasoning tokens in your logs.

Play 5: Route each task to the effort it needs

Goal: use one setting per job, not one setting for the whole app. Anthropic’s best-practice list calls this dynamic effort.

  1. Map each label from Play 1 to the level your sweep chose.
  2. Send the effort with each request based on the job type.
  3. Keep the level steady within one long conversation if you rely on prompt caching; changing it can break the cache.
  4. Re-run the sweep when you change model. Levels are tuned differently on each model, so old settings may not carry over.

Meera’s final setup: low for order-status questions, medium for refund disputes, and high only for a weekly report her manager reads.

You are done when: each job type sends its own effort level, and you have a note to re-test after model changes.

Do you need a reasoning model at all?

The table below sums up reasoning models for beginners in one view.

ChoiceGood forTrade-off
No or minimal thinkingLabels, lookups, voice, reformattingFastest and cheapest, weak on multi-step logic
Low effortHigh-volume chat, short tool tasksCheap; may skip checks on long tasks
Medium effortMost apps, agents, coding helpGood balance; the usual starting point
High effortHard logic, debugging, many rulesSlower, more output tokens
Highest levels (xhigh, max)Rare, very hard or long tasksMuch more cost; use only when tests show a gain

If you are still choosing a model, read how to choose an AI model first. Sometimes a smaller model at higher effort beats a bigger one at low effort, and only a test tells you.

Mistakes to avoid

  • Setting the highest effort everywhere. Fix: run a sweep and use the lowest level that passes.
  • Using effort to fix a vague prompt. Fix: write the goal, rules and success check first (Play 2).
  • Forgetting the hidden tokens. Fix: raise the output limit and log reasoning tokens.
  • Expecting shorter replies from lower effort. Fix: Anthropic notes effort barely changes visible reply length on some models; ask for length in the prompt.
  • Copying settings to a new model. Fix: re-run your sweep after every model change.

How to know it is working

  • Your test score at the chosen level is within a point or two of the highest level.
  • Average wait time and cost per request went down, or stayed flat while quality rose.
  • No answers are cut off or empty in your logs.
  • Hard tasks get more effort than easy ones, and you can see it in the request logs.

Your checklist

  • ☐ Every AI job labelled easy, medium or hard
  • ☐ Prompts for hard jobs have a goal, rules and a success check
  • ☐ Effort sweep done at low, medium and high
  • ☐ Output limit leaves room for thinking
  • ☐ Reasoning tokens logged per request
  • ☐ Effort set per job type, re-test planned for model changes

Frequently asked questions

Is a reasoning model always better?

No, and this is the first rule of reasoning models for beginners. On easy tasks it gives about the same answer as a normal model, but slower and at higher cost. It shines on multi-step problems with a clear right answer.

Can I see the model’s thinking?

Usually not in full. OpenAI keeps reasoning tokens hidden and may offer a summary; Anthropic can return thinking blocks on supported models. Either way, you pay for the tokens.

What effort level should a beginner start with?

Medium, then test low and high on your own questions. Both OpenAI and Anthropic suggest measuring before you move to the highest levels.

Does higher effort make answers longer?

Not reliably. Effort mostly changes how much hidden thinking happens. If you want shorter or longer answers, say so in the prompt.

Is reasoning effort the same as temperature?

No. Temperature controls how random the word choices are. Effort controls how much the model plans and checks before answering. They are separate settings.

Back to Meera’s bot

A week later, Meera’s bot answers order questions at low effort in about a second, and refund disputes at medium with far fewer mistakes. The bill went back below where it started (illustrative), and her teammate now runs a sweep before touching any dial. That is the whole idea of reasoning models for beginners: let the AI think only as hard as the job needs.

Your next step: take your five most common AI requests, label them, and run one small sweep this week. Then pair it with choosing the right model, because model and effort work best when you tune them together.

Sources: Anthropic: Effort (levels, defaults, behavioural signal, best practices); OpenAI: Reasoning models (reasoning tokens, billing, effort values, reserving tokens).

1 comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Newsletter

Get the next guide before anyone else.

One email a week with new playbooks and tutorials. No spam, unsubscribe anytime.