Reasoning Models: When to Use AI That Thinks Longer
How reasoning models differ, which tasks are worth the wait, and how to prompt models that think for themselves, based on OpenAI and Anthropic guidance.
You switch on "think longer" to have AI write 20 product captions. You wait several times longer and the result is barely different. Another day you use a fast model to compare 3 supplier contracts, and it misses a penalty clause buried in one of them. Neither case means the AI is bad. It means the wrong model was picked for the job.
What is a reasoning model?
A reasoning model spends time "thinking" internally before answering: working through logic, checking itself or planning several steps. Providers name this differently. OpenAI has its o-series models, and Anthropic calls this capability in Claude extended thinking, with adaptive thinking that adjusts how much it thinks.
OpenAI's reasoning best practices guide draws a clear comparison: reasoning models are like "planners" that excel at strategy, multistep problem-solving and decisions based on ambiguous information, while general models are "workhorses" that execute well-defined tasks quickly and at lower cost. Many workflows use both.
Which tasks are worth a reasoning model?
OpenAI lists these good fits:
- Ambiguous tasks: inferring intent from limited or scattered information, often asking clarifying questions.
- Needle in a haystack: pulling the key facts out of large volumes of text.
- Complex documents: finding relationships and nuance across contracts, financial statements or insurance claims.
- Multistep planning: breaking down a problem and assigning each step to the right model.
- Code review and debugging: catching subtle issues spread across files.
- Evaluating other models' outputs: acting as a grader.
Tasks where a fast model is enough
Well-defined, high-volume or instant-response work, such as classifying messages, translating short text, summarising emails or writing captions from a template. OpenAI notes that general models are faster and tend to cost less, while reasoning models suit cases where accuracy matters more than speed and cost.
How to prompt a reasoning model
Short, direct, with success criteria
OpenAI recommends simple prompts, explicit constraints and success criteria, and delimiters such as headings or XML tags to separate parts of the input.
No need to say "think step by step"
OpenAI notes that these models already reason internally, so chain-of-thought instructions are unnecessary. Anthropic recommends general instructions, such as asking it to think thoroughly, rather than writing a step-by-step plan, because the model's reasoning often goes further than a plan a person would prescribe.
Try zero-shot first
OpenAI recommends starting without examples and adding them only if needed (see Few-shot Prompting).
Ask it to check its answer
Anthropic recommends ending with an instruction to verify the answer against test criteria before finishing. It helps catch errors, especially in coding and maths.
Example: an SME choosing a supplier
<contracts> three supplier contracts </contracts> Goal: recommend the lowest-risk supplier for a shop with seasonal sales swings. Criteria: compare minimum order terms, penalties and cancellation terms. Quote the contract text for every conclusion. Before answering, check that every conclusion has a real quote behind it.
This task requires reading across documents and weighing trade-offs, so a reasoning model is worth it. Writing the email to notify suppliers after the decision only needs a fast model.
Not every task needs a model that thinks longer, but every task deserves the model that fits it.
Things to watch out for
- Thinking longer does not mean always right. Still check important facts.
- Cost and waiting time are higher. Test whether it pays off before using it at volume.
- For legal and contract matters, AI can help with a first analysis, but consult a professional before making real decisions.
How to start
- List the AI tasks you use and split them into well-defined tasks and tasks that need judgement.
- Take one judgement task and run it on both a fast model and a reasoning model with the same input.
- Compare quality, time and cost, and decide based on real results.
- Shorten the prompt, add success criteria and ask the model to check its own answer.
Knowing which kind of AI to use when is becoming a skill employers look for. Talk to the AI Career Assistant to plan your AI skill development, or browse jobs on Plants.work.