Every major AI model now has a setting for how hard it thinks before it answers. Most people never touch it. They run a pricing decision on the default, or they leave the top setting on for jobs a fast model finishes in seconds. The first mistake costs you accuracy. The second costs you money and waiting time.
Two years ago this was a choice between products. OpenAI sold o1 as a separate reasoning model, and you picked it when a job was hard. That product line is being shut down: o1-pro, o3-mini and o4-mini were switched off on 23 October 2026, and o3 goes on 11 December. Reasoning now lives inside the main models as a setting. Anthropic calls it effort, OpenAI and Grok call it reasoning effort, Google calls it the thinking level, and DeepSeek has a thinking mode. In the chat apps it’s a menu or a slider next to the model name. Through the API it’s a parameter you set on each request.
So the question has changed. You no longer ask which reasoning model to use. You decide how much thinking each job deserves.
1. What reasoning means here
By reasoning I mean the model working through intermediate steps before it gives you an answer.
A factual question like “What is the capital of the USA?” needs none. A question like this one needs a lot:
“If our 7-day onboarding email sequence is converting new subscribers into sales leads at 10%, what can we do to increase the conversion rate and shorten the time from new subscriber to warm lead?”
To answer it, the model needs your data and your emails, and it has to work through several possible causes before it recommends anything. At a high setting that can take minutes. On this kind of question, that’s time well spent.
The gap between thinking and not thinking was easy to see in the first generation of these models. On the 2024 AIME, a US math competition, OpenAI reported GPT-4o solving 12% of problems and its first reasoning model, o1, averaging 74%. DeepSeek reported 79.8% for R1 on the same exam. Those were separate models then. Today the same gap sits between the low and high settings of one model.
2. What more thinking costs
You pay for the thinking. Anthropic, OpenAI and Google all bill the tokens a model spends reasoning as output tokens, even when you never see them, and Grok’s API charges them at the full output price. At the extreme the bill is enormous. ARC Prize reported in December 2024 that a preview of o3, at its highest compute setting, used 5.7 billion tokens across 100 ARC-AGI puzzles, about 57 million each, and took 13.8 minutes per puzzle. OpenAI has since released a different version of the model.
Ordinary business jobs cost far less, but the spread between settings and models is still wide. On Artificial Analysis’s leaderboard in early November, GPT-6.1 Sol at its top setting scored 55 at $1.27 a task, while Claude Opus 5.5 at its top setting scored 59 at $10.04. I keep current prices and settings in which AI model for which job.
Higher settings also take longer and write more. On a simple job, that’s all the extra spend buys you.
3. My routing rule
My rule is simple. If the task touches pricing, revenue or strategy, turn the setting up. If it’s repetitive writing or extraction, use a fast model on its default or lowest setting. If it doesn’t need AI at all, don’t spend anything.
Before I give a workflow a high setting, I run it through four questions:
- Does the task need several steps of deduction? If not, use a low setting or skip AI.
- Is the cost under 5% of the potential revenue impact? If you can’t justify the spend, keep the setting down.
- Can a lower setting get 90% of the quality at 10% of the cost? Run both on five real examples from your business before you decide.
- Is there a separate check or a human review? If not, the model’s confidence is a liability.
Here is what happens when you skip this. At 2 a.m., a B2B SaaS founder I worked with pulled a competitive pricing analysis from a fast model on its default setting and felt relief. The output looked polished, the structure was clean, and the answer sounded confident enough to drop into a board deck before sunrise.
By noon, two competitor numbers were wrong, and the board saw the whole AI initiative differently. Nobody cared that the prose was elegant. They cared that the machine had acted certain while being careless.
What he had was a routing problem. A job that needed deliberate inference went to a setting built for speed, and it cost him the board’s trust. If a mistake would travel into pricing, forecasting or strategy, don’t trust a single fluent pass.
Here is how that rule maps onto the work most small businesses hand to AI.
| The job | Setting | What I’d use |
|---|---|---|
| Tagging leads, summarising call notes, first drafts nobody sends | Fast model, lowest or default setting | Claude Haiku 5.5, GPT-6 Luna, Gemini 3.5 Flash-Lite |
| Proposals and client deliverables you read before they go out | Mid-tier model, default setting | Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash |
| Pricing, offer positioning, decisions with money on them | Top model, high setting, plus a separate check | Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon |
These names change every few months, so treat the table as a starting point. The current list and prices are in which AI model for which job.
4. Where to find the setting, and what it’s set to now
In Claude, click the model name next to the send button, then Effort, and pick from Low to Max. In ChatGPT, the Thinking slider runs from Instant through Medium and High to Extra High and Pro, though Plus stops at High. In the Gemini app, pick a model and then a thinking level of Low, Medium or High. Grok’s app has its own reasoning modes.
Through the APIs, Claude Opus 5.5 and Sonnet 5.5 take an effort setting, GPT-6.1 Sol and GPT-6 Astra take a reasoning effort, Gemini 4 Argon takes a thinking level, Grok takes a reasoning effort, and DeepSeek V4 Pro has a thinking mode with its own effort setting.
Two things in the vendors’ documentation surprised me. First, on several of these models you can’t switch reasoning off. Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Grok all think on every request, so a lower setting is the only way to make them think less. Second, the defaults differ, even within one company. Through the API, Claude Opus 5.5 defaults to medium and Sonnet 5.5 to high, and Grok defaults to high. The Gemini app defaults to Low unless you choose otherwise. If you’ve never checked, you don’t know what you’re paying for on any given job.
If you want one interface for several of these, Poe gives you models from each of these companies in one place, and each one keeps its maker’s own setting: Effort on Claude, Reasoning effort on GPT and Grok, Thinking level on Gemini.
There are smaller models that reason too, and more specialised ones will keep arriving. If you understand when to turn thinking up, a new model is just one more option to test.
5. What a model does when it reasons
Researchers break reasoning into a few skills. You use the same skills yourself, and knowing them helps you write better instructions.
Inference
Inference means drawing new conclusions from the information at hand, instead of retrieving something already stated. Given “All dogs can bark, and this animal is a dog,” a model with good inference concludes “this animal can bark.”
Source: Natural Language Reasoning, A Survey
Problem-solving
When a request is unclear or contains conflicting details, a model reasoning well considers other readings and adjusts. That matters on questions with no straightforward answer, where it has to weigh pros and cons and bring in wider context. The Tree-of-Thought approach is one way researchers structured this: the model explores several branches and backs out of dead ends.
Source: Large Language Model Guided Tree-of-Thought
Understanding context
To answer the onboarding question above, you’d want email metrics, the emails themselves, profiles of new subscribers and warm leads, and notes from sales conversations. A model needs the same.
This is why so many prompts fail. They assume the model knows your business. It doesn’t. Without your data it will still answer, but the answer will be generic and often wrong. Turning the setting up on a thin prompt just gets you a longer version of the same wrong direction.
Handling complexity
Complex reasoning means multi-step deduction, keeping track of details and weighing conflicting information. A business question with clues spread across several documents is a good fit: the model has to read each piece correctly and work out how they relate. The paper Large Language Models as an Indirect Reasoner walks through this in more detail.
Switching between fast and slow
Researchers built frameworks that switch between fast, intuitive answers and slow, deliberate reasoning depending on how hard the problem is. That idea is now in the products. Claude’s current models use what Anthropic calls adaptive thinking, deciding for themselves how much to think within the effort level you set. Gemini models adjust their thinking to the difficulty of each request by default.
Source: HDFlow: Enhancing LLM Complex Problem-Solving with Hybrid Thinking and Dynamic Workflows
You do this yourself. On most problems there’s an intuitive first pass, then slower checking, and you move back and forth between them.
6. The five types of reasoning in your business
Reasoning is usually split into types. This illustration contrasts two of them, deductive and inductive.
Source: Inductive or Deductive? Rethinking the Fundamental Reasoning Abilities of LLMs
Deductive reasoning
Deduction applies a general rule you already have to a specific case. An ecommerce company has a policy that any product under a 10% profit margin gets discontinued. Product X runs at 7%, so under the policy it goes.
That’s a job you can hand to an agent: scan sales data every week, apply your rules, and flag what to cut, what to keep and which products are drifting toward the line. The rules are yours. The model applies them consistently.
Inductive reasoning
Induction draws a general conclusion from observed patterns. It suits messy data where the rules aren’t written down, and its conclusions are probable, never certain.
Say an email team notices that its last five campaigns with emoji subject lines had 25% to 30% higher open rates, and makes emojis a rule. Five campaigns is a small sample, and Apple’s privacy feature inflates open rates, so this is exactly the kind of conclusion to test before it becomes policy. A model reasoning well should say so. If yours doesn’t, ask it what would make the conclusion wrong.
Abductive reasoning
Abduction looks for the most likely explanation when the information is incomplete. A software company sees a sudden 30% drop in app engagement. The most likely cause is a bug in the latest update that makes features harder to reach. A competitor launch or a seasonal dip are possible too, but the bug is the best first bet, so that’s where the fixing starts.
This is where a high setting earns its cost: the model can list the candidate explanations, say what evidence would separate them and suggest what to check first.
Analogical reasoning
Analogy carries what’s known about one situation to another that looks different on the surface. A startup struggling with customer acquisition notices that its position resembles streaming services before they added free tiers, and tests a free tier with paid upgrades.
Give a model a description of your problem and ask which businesses in other industries faced the same structure, and what happened to them. It gets you past “we’re different” thinking faster than a brainstorm.
Practical reasoning
Practical reasoning recommends an action from your goals, your constraints and the likely outcomes. A manufacturer weighing new automation sets out its goals (more output, lower costs), constraints (limited budget, training) and likely outcomes (better efficiency, early production delays), then decides to roll it out in phases.
A model on a high setting is a useful partner for this kind of decision, provided you give it the real constraints and check its reasoning against what you know.
7. Where each kind of work sits
Content and copy
Content planning across seasonal trends, audience data and business goals benefits from a higher setting, because the model is weighing several inputs. First drafts from a clear brief don’t need one. Editing does: ask the model to find where an argument contradicts itself or loses the reader, which is abductive work.
Ads and campaigns
Generating ad variants from a product sheet and audience profiles is fast-model work. Analysing why one ad set underperformed is not. That’s a search for the most likely explanation, and it deserves the higher setting and your performance data.
Workflow automation
Support triage and pulling tasks and deadlines out of email can run on a fast model, with an escalation path to a higher setting for unusual cases. Decisions inside an automation, such as whether to reorder stock now or wait given seasonal history, need the higher setting and rules you’ve written down.
Data analysis and decisions
Finding patterns in raw data, explaining a sudden change and comparing a decision to similar past situations all reward more thinking. Turning finished numbers into a report usually doesn’t.
8. Three patterns I use in practice
Here is how I wire reasoning into a workflow. You don’t need a platform for any of this, and you can build it with any major model API.
The first is planner then executor. A model on a high setting breaks the goal into tasks, and a fast model carries them out. I use this when execution speed matters but the plan has to be right.
The second is verifier after draft. A fast model writes the first pass, then a model on a high setting audits it before anything ships. That’s useful for landing pages, pricing pages and campaign copy, where a hidden contradiction does real damage.
The third is a tool loop. The model alternates between thinking and calling tools until it reaches a result. This is the pattern for agents, research workflows and anything that needs live data. Current models can think between tool calls, which is what makes the loop work (OpenAI’s reasoning guide and Anthropic’s thinking docs both describe it).
Put together, the architecture is simpler than most people expect. An orchestrator receives the job and picks the path. A model on a high setting handles decomposition, verification and decisions. A fast model writes, formats or extracts once the plan is clear. A verifier checks for contradictions, missing steps or broken logic. A tool layer pulls documents, CRM records, ad data or analytics. I write up workflows like these in Bionic Business.
When one of these setups fails, I usually find the problem in the context or the routing. The model can only reason over what you give it, which is why I keep pointing people to context engineering vs prompt engineering.
9. Treat visible reasoning as a clue
Visible reasoning feels reassuring. It shouldn’t. Anthropic has shown that reasoning models don’t always say what they think, so the explanation you see can be incomplete or misleading (Anthropic’s research on reasoning models).
What you get is a generated explanation, with no guarantee that it matches what happened inside the model. The output can look careful and still be wrong. I use three rules in production:
- Verify the conclusion separately. Don’t let the model grade its own homework.
- Log inputs as well as outputs. If an answer fails, you need to see the prompt, the context it was given and the tool calls.
- Put a person on high-stakes decisions. Pricing, legal, customer risk and anything that commits revenue needs human review.
10. What to change this week
Start by checking what setting each tool you use is on. Then write down every job you currently hand to AI. Mark the ones where a mistake would reach a price, a forecast, a client or a strategic decision. Those get the higher setting and a separate check. Everything else starts on a fast model at its default.
Then take one job from each group and run five real examples at two settings, side by side. Keep the lowest setting that still gets the answer right, and write that setting into the workflow so nobody has to remember it.
Sources
- OpenAI, Learning to reason with LLMs, September 2024 (AIME 2024 results for GPT-4o and o1)
- DeepSeek-AI, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, January 2025
- ARC Prize, OpenAI o3 breakthrough high score on ARC-AGI-Pub, December 2024
- Artificial Analysis, model leaderboard, read early November 2026
- Yu et al., Natural Language Reasoning, A Survey, 2023
- Long, Large Language Model Guided Tree-of-Thought, 2023
- Zhang et al., Large Language Models as an Indirect Reasoner: Contrapositive and Contradiction for Automated Reasoning, 2024
- Yao et al., HDFlow: Enhancing LLM Complex Problem-Solving with Hybrid Thinking and Dynamic Workflows, 2024
- Cheng et al., Inductive or Deductive? Rethinking the Fundamental Reasoning Abilities of LLMs, 2024
- Anthropic, Reasoning models don’t always say what they think, 2025
- OpenAI, Reasoning models guide, deprecations and GPT-6 in ChatGPT, read 7 November 2026
- Anthropic, effort, thinking and model, effort and thinking settings in Claude, read 7 November 2026
- Google, Gemini thinking and About Gemini models, read 7 November 2026
- Grok API, reasoning, read 7 November 2026
- DeepSeek, thinking mode, read 7 November 2026
