Before you start
You need about ten real tasks from your own work, a reasoning model and a fast model from the same provider. A planner is the vendor's word for a reasoning model that decides what to do. A workhorse is a faster model that carries the plan out.
Sort your tasks and test both models
A small test answers the question for your work better than any benchmark can. Ten tasks are enough to show a pattern.
- List ten tasks you give to an AI model today. For each, write what a finished answer looks like and how you would check it.
- Mark each task as planning or as execution. Planning means the answer depends on a chain of earlier steps, as in maths, debugging or a decision from messy documents. Execution means the job is well defined, as in extracting fields, reformatting, routing or classifying.
- Note which matters more for the task, speed and cost or accuracy and reliability.
- Run every task on the reasoning model and on the fast model, and mark each reply pass or fail against your check.
- Keep the fast model for every task where both pass. Compare the models on cost per finished task, not on price per token.
Where a reasoning model earns its time
OpenAI's best practices page calls its reasoning models the planners and its GPT models the workhorses. The planners think longer on complex tasks, work through ambiguity and suit jobs that would otherwise need a human expert. The workhorses are the lower-latency, cheaper models built for straightforward execution.
The page lists where customers saw the planners pay off, from a vague request with gaps to fill to one needle in a haystack of documents. The rest are relationships across hundreds of pages, multistep planning for an agent, visual reasoning, reviewing and debugging code, and grading other models' answers. Most workflows, it says, use both kinds. The reasoning model plans and the GPT model executes.
Its own example is a returns desk. GPT-4o and GPT-4o mini triage the order details and identify the issues and the return policy. They feed all of it to o3-mini, which makes the final call on whether the return is viable.
The matrix turns those sources into a rule of thumb. It is our reading of the vendors' advice, not a measurement. OpenAI's effort table lists classification at its none level and customer support and chat assistants at low, and Anthropic says chat and classification often do well at low effort. A reasoning model at a low setting can therefore serve in the last row. Reasoning effort covers those settings.
| Reasoning model | Fast model | |
|---|---|---|
| A vague request with gaps to fill | Yes | Partly |
| One clause hidden in dozens of documents | Yes | Partly |
| Planning the steps for an agent | Yes | Partly |
| Reviewing code and grading other models' answers | Yes | Partly |
| Well-defined execution where speed and cost matter | No | Yes |
| Chat, classification and other latency-sensitive replies | Partly | Yes |
Sources: OpenAI Docs: Reasoning best practices (read November 6, 2026), OpenAI Docs: Reasoning models (read November 6, 2026), Anthropic on GitHub: claude-api skill (read November 6, 2026)
What customers told OpenAI
The same page backs its list with customer quotes about o1. Blue J, SafetyKit and a Braintrust customer compare it with GPT-4o, and Hebbia compares it with other models. OpenAI picked the quotes. The models are old. Read them as the kinds of task where reasoning helped, not as today's gap.
88%
SafetyKit's accuracy on its hardest image classification tasks with o1, against 50% with GPT-4o
Source: OpenAI Docs: Reasoning best practices (read November 6, 2026)
0.74
F1 score of a Braintrust customer's judge model with o1, against 0.12 with GPT-4o
Source: OpenAI Docs: Reasoning best practices (read November 6, 2026)
4x
Blue J's end-to-end improvement on tax research after swapping GPT-4o for o1
Source: OpenAI Docs: Reasoning best practices (read November 6, 2026)
52%
of complex Credit Agreement prompts where Hebbia found o1 stronger than other models
Source: OpenAI Docs: Reasoning best practices (read November 6, 2026)
Where a reasoning model fell behind
DeepSeek's R1 model card shows the other side, because it reports DeepSeek-V3 and the reasoning model R1 on the same benchmark rows. Of the 20 rows where the card gives both scores, R1 was higher on 18. It was lower on IF-Eval, an instruction-following test, and on C-SimpleQA, the Chinese version of a short factual question test.
The drops are small, 2.8 and 4.3 points. Both models come from one company. They still show that training a model to reason does not make it a strict upgrade.
Anthropic says something similar from the other direction. Its Opus 4.6 announcement says the model thinks more deeply on harder problems, which can add cost and latency on simpler ones, and recommends dialing effort from high to medium if it overthinks. Anthropic's claude-api skill adds that coding and long-horizon agent work respond strongly to more effort, while chat, classification and high-volume or latency-sensitive routes often do well at low.
Source: Hugging Face: DeepSeek-R1 model card (read November 6, 2026)
Research on where extra thinking stops helping
Three papers from 2024 and 2025 went looking for the limits. Their abstracts name model families more than versions: OpenAI o1 in the first, Claude and OpenAI o-series models in the second with one finding about Claude Sonnet 4, and no model in the third. Use them as failure modes to watch for and not as a ranking. The third also saw reasoning effort rise with difficulty and then fall, even with token budget left.
| Paper | What it looked at | What it found |
|---|---|---|
| Do NOT Think That Much for 2+3=? (December 2024) | o1-like models on problems of different difficulty | They spend excessive compute on simple problems for minimal benefit. The authors propose training that cuts it without losing accuracy |
| Inverse Scaling in Test-Time Compute (July 2025) | Tasks built so that longer reasoning hurts: counting with distractors, regression with spurious features, deduction with constraint tracking | Accuracy fell as reasoning grew longer. Claude models became more distracted by irrelevant information, and OpenAI o-series models overfit to how a problem was framed |
| The Illusion of Thinking (June 2025) | Puzzles whose complexity can be dialed up | Standard models beat reasoning models on low-complexity puzzles, reasoning models won at medium complexity, and both collapsed at high complexity |
A video that covers the same question
Andrej Karpathy's long walkthrough of how he uses LLMs has a chapter called "Thinking models and when to use them", which starts at 22:54 in the video's chapter list. He posted it on 27 February 2025, so the models it names are older than the ones in this course.
Common mistakes
- Choosing a reasoning model for everything because it tops benchmarks. OpenAI treats the two families as different tools, not as a ranking.
- Comparing price per token. Anthropic's cost guide says to optimize cost per completed task, since a model with a higher price can be cheaper if it finishes in fewer turns.
- Trusting a hard answer without a check. In The Illusion of Thinking, accuracy collapsed completely once the puzzles got hard enough.
- Reusing an old test. Anthropic's guide says the effort curve is per workload and per model, so re-sweep after a model migration, a major prompt change or a workload shift.
Try it yourself
- Run the ten-task test above and count the tasks where only the reasoning model passed.
- Find one job in your work that a planner could plan and a cheaper model could carry out, and write the two prompts.
- Save the results with today's date, and run the same ten tasks again after your provider's next model release.
Next course: Prompt optimization
You now have a way to choose a model for each task. The next course teaches AI to draft, improve and test the prompts themselves, and it opens with AI prompt generators before it moves on to testing prompts.
Check yourself
In OpenAI's split, which job suits the "workhorse" GPT models?
Show answer
B. Straightforward execution where speed and cost matter OpenAI describes the workhorses as the lower-latency, cheaper models designed for straightforward execution.
On which benchmark did DeepSeek's model card show R1 below V3?
Show answer
C. IF-Eval The card lists 86.1 for V3 and 83.3 for R1 on IF-Eval, one of two rows where R1 was lower.
What does Anthropic's cost guide say to compare when you choose between models?
Show answer
D. Cost per completed task The guide says to optimize cost per completed task, because a pricier model can finish in fewer turns and cost less overall.
In the Illusion of Thinking study, which tasks did standard models handle better than reasoning models?
Show answer
A. Low-complexity tasks The paper found three regimes: standard models won at low complexity, reasoning models at medium, and both collapsed at high.
Frequently asked questions
When should you use a reasoning model instead of a regular one?
When the task has steps that depend on each other and a wrong answer costs more than a slow one. OpenAI's examples include reviewing code, finding one clause across dozens of documents, planning for an agent and grading other models' answers.
Are reasoning models better than regular models?
Not across the board. OpenAI says neither family is better, only different, and DeepSeek's own table has R1 behind V3 on two of 20 rows. Among that table's scores on a 0 to 100 scale, R1's biggest jumps came on contest-style tests, about 40 points on AIME 2024, 38 on the Codeforces percentile and 36 on CNMO 2024.
Why is my reasoning model slower and more expensive?
Because it writes thinking before it answers, and OpenAI bills those tokens as output. OpenAI describes its GPT models as faster and tending to cost less. If a task needs a reasoning model but not a deep one, reasoning effort cuts the thinking.
Do reasoning models still get things wrong?
Yes. The Illusion of Thinking study found a complete accuracy collapse beyond certain complexities in its puzzle environments, so check hard answers with tests, a validator or a second source.
Is a reasoning model better for coding?
It can be for reviewing and improving large amounts of code. OpenAI says reasoning models are particularly effective at that, often running in the background because of their higher latency, and Anthropic lists coding among the workloads that repay higher effort. For writing code where latency matters, OpenAI says its GPT models may be better designed.
Sources
- OpenAI Docs: Reasoning best practices https://developers.openai.com/api/docs/guides/reasoning-best-practices
- OpenAI Docs: Reasoning models https://developers.openai.com/api/docs/guides/reasoning
- Hugging Face: DeepSeek-R1 model card https://huggingface.co/deepseek-ai/DeepSeek-R1
- Anthropic: Introducing Claude Opus 4.6 https://www.anthropic.com/news/claude-opus-4-6
- Anthropic on GitHub: claude-api skill https://github.com/anthropics/skills/blob/main/skills/claude-api/SKILL.md
- Anthropic on GitHub: claude-api skill, cost optimization guide https://github.com/anthropics/skills/blob/main/skills/claude-api/shared/cost-optimization.md
- Hugging Face Papers: Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs (2024) https://huggingface.co/papers/2412.21187
- Hugging Face Papers: Inverse Scaling in Test-Time Compute (2025) https://huggingface.co/papers/2507.14417
- Hugging Face Papers: The Illusion of Thinking (2025) https://huggingface.co/papers/2506.06941
- YouTube: Andrej Karpathy, How I use LLMs https://www.youtube.com/watch?v=EWvNQjAaOHw