Reasoning models, when to make them think
Reasoning models "think" before answering. How they work, what effort levels are, what they cost in time and money, and when they are really worth it.
Mattia PanattaUpdated 15 October 202610 minIntermediate
In short
- A reasoning model writes a long internal reflection before the answer: it tries paths, corrects itself, checks.
- The more it “thinks”, the better it does on hard problems, but the more it costs and the slower it is. From low to max effort, GPT-6.1 Sol goes from 15 seconds to almost 5 minutes per task.
- Use it for analysis, maths, code, decisions with many constraints. For emails, summaries and translations it’s waste.
- With these models, describe the goal well and leave the method to them.
Thinking out loud, silently
Until 2024 chatbots answered “off the cuff”: token after token, without pausing. Then came a simple, powerful discovery: if you have the model write out its reasoning before the answer, results on hard problems improve enormously. Each reasoning token is one more small step of computation.
The first models of this kind arrived in late 2024, and in January 2025 DeepSeek R1 showed the technique worked on open, cheap models too. Today practically every top model reasons: Claude Opus 5.5, GPT-6, Gemini 4 Argon, Muse Wonder.
How do they learn to reason? Through reinforcement on verifiable problems: the model gets millions of maths, logic and programming problems with checkable answers. It tries, is rewarded when right, and over time develops useful strategies on its own: breaking the problem down, checking steps, backtracking when something doesn’t add up.
Effort: a dial between quality, time and cost
Many models let you choose how much to reason: levels like low, medium, high, xhigh, max. In apps the system often chooses automatically, or you set it with options like “Think longer”.
Artificial Analysis measures the same model at different levels. Here is GPT-6.1 Sol (the model behind Lumen’s chatbot):
Reading the numbers:
- from low to medium the index rises 6 points, with a modest cost increase and the same time (15 seconds): almost always a good deal;
- from medium to max you gain another 4 points, but cost per task triples and time goes from 15 seconds to almost 5 minutes.
The same pattern holds for everyone: Claude Opus 5.5 goes from 42 points (low, $0.55 per task) to 58 (max, $5.98).
The hidden cost
Reasoning tokens are billed as output tokens, even when you don’t see them. A 200-word answer may be preceded by 5,000 tokens of thinking. Via API, it’s the line item that inflates the bill most.
When you need it (and when you don’t)
| Task | Reasoning | Why |
|---|---|---|
| Emails, posts, translations, summaries | None or low | Needs fluency, not deduction |
| Extracting data from a document | Low | Reading well matters, not thinking long |
| Analysing a contract or financial statements | Medium or high | Many elements to connect and check |
| Calculations, financial models, statistics | High, with code execution | Every step must be verified |
| Writing a complex function, debugging | High | Needs to test hypotheses |
| Planning a project with many constraints | High | Scenarios, alternatives, dependencies |
| Open research problems | Max | The extra time pays off |
A rule of thumb: if you’d need pen and paper, the model needs to reason. If you’d answer off the cuff, a fast model is enough.
How to prompt a reasoning model
With “classic” models the advice was to guide them step by step (“first do A, then B, then C”, “think step by step”). With reasoning models that’s often counter-productive: they’ve learned better strategies than ours. What works better:
- Describe the goal and expected result, not the procedure.
- Give all constraints and data in one go, even if the message gets long.
- Say how to verify: “check the totals add up”, “make sure the code passes these tests”.
- Ask for a clear final format, so you don’t get the reasoning instead of the answer.
Goal: decide whether my client should move from the flat-rate regime to the ordinary regime in 2027.
Data: expected revenue €78,000, actual costs €31,000, profitability coefficient 78%, no employees, separate-management pension contributions.
Constraints: consider only taxes and contributions, not VAT to customers. State your assumptions.
Verification: recompute the totals twice and flag any missing data.
Format: a comparison table of the two regimes, then a 3-line recommendation and a list of points to check with a professional.
What’s new in 2026: reasoning and answering together
Waiting is the main drawback of these models. GPT-6, rolled out in ChatGPT in October 2026, tries to reduce it: it starts answering while still thinking, interleaving reasoning with partial answers. According to OpenAI, GPT-6 at Extra High starts answering as quickly as the previous model at medium, with better results. Other labs are heading the same way: models that decide on their own how long to think based on how hard the question is.
In practice
- In apps, leave automatic selection on for normal questions and turn on extended reasoning only for analysis, maths and complex problems.
- Via API, start at medium: the best compromise in most cases. Go higher only if you measure an improvement on your task.
- Don’t ask models that already reason to “think step by step”: ask how to verify the result instead.
- If you’re in a hurry and the task is simple, a fast model like Claude Haiku 5.5 (240 tokens per second) or GPT-6 Luna beats any thinking model.
Sources
Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.
