Making ChatGPT Reason: From Chain-of-Thought to Reasoning Levels

For a long time, the so-called chain-of-thought technique was the most well-known trick for improving answers on difficult problems: adding “reason step by step” to the end of the prompt. The models used in ChatGPT today reason internally, and OpenAI’s documentation has shifted its stance. This lesson explains what no longer works, what remains useful, and how to get a thoughtful answer that you can verify.

What Has Changed

OpenAI’s guide on reasoning models is explicit: these models think before they answer, so asking them to “think step by step” or “explain their reasoning” is useless, and some techniques of this type can even degrade the result. Instead, it recommends simple, direct prompts.

In ChatGPT, you control the depth of reasoning with the model selector: in Chat mode, from Instant to Extra High, then Pro on plans that offer it; in ChatGPT Work and Codex, with the power setting. When a problem resists, the right answer is no longer to add a magic formula, but to raise the reasoning level. The course OpenAI Reasoning on this site details these settings.

What Remains Useful: Framing the Problem Well

A reasoning model needs substance and a clear objective. OpenAI’s guide advises being very specific about the expected outcome, stating constraints explicitly, and separating parts of the prompt. In practice, four elements make the difference: all useful data, the assumptions to keep or flag, your success criteria, and the response format.

Take a profitability question. Instead of writing “reason step by step,” ask for an answer you can control line by line:

Our revenue is €2.4M, our fixed costs are €1.8M,
and our variable costs represent 25% of revenue.
Are we profitable?

Present the calculation in a table (variable costs, total costs,
profit), then conclude in one sentence.

The table isn’t there to make the model think: it’s there to let you verify. Here, you should see variable costs of €0.6M, total costs of €2.4M, and zero profit: the company is exactly at the break-even point. If a figure differs, you see immediately where.

Asking for assumptions follows the same logic. “State the assumptions you are making” brings implicit postulates to light, such as an assumed constant growth rate, which can skew a conclusion without ever appearing in it. “Start by restating the problem in one sentence” gives you a control point before the model produces three pages on the wrong question.

Structuring the Expected Response

Structure remains valuable, provided it focuses on what you want to receive, not on how to think. For a decision, it forces a comparison before a verdict and integrates costs that a quick comparison misses: recruitment, ramp-up time, turnover risk.

I need to choose between hiring a senior developer (€75,000 per year)
or outsourcing development to an agency (estimated €90,000 per year).
Our backlog represents 8 months of work.

Structure your response as follows:
1. Advantages and risks of each option
2. Real cost over 24 months, including recruitment and onboarding
3. Non-financial factors
4. Argued recommendation, with the assumptions made

For a diagnosis, demand a wide range of causes before ranking: without the “at least 8,” the response risks stopping at the first plausible explanation.

Our conversion rate dropped from 3.2% to 1.8% in two months.
Traffic is stable, the product hasn't changed, but we redesigned
the page 6 weeks ago.

1. List at least 8 possible causes
2. Rank them by probability
3. For the 3 most likely, indicate how to verify them
4. Propose a prioritized action plan

For a critical evaluation, a grid of specific points avoids complacent commentary. If the grid asks for a market comparison, require web research and sources: without them, the cited “market standards” risk being invented.

Here is our SaaS pricing grid: Starter €29 per month,
Pro €79 per month, Enterprise upon request.

Evaluate it based on these points:
1. Consistency of the gap between plans
2. Gaps in the range
3. Comparison with three competing offers, citing your sources
4. Risks (cannibalization, barrier to entry)
5. Proposed adjustments, each justified

Reasoning Backward from the Goal

To plan, start from the goal and work backward to the means. This is a quick way to turn a numbered ambition into a credible plan, or to discover that the goal requires a budget you don’t have.

Our goal: €100,000 in monthly recurring revenue in 12 months.
We are at €35,000. Average basket: €200 per month.
Lead-to-customer conversion rate: 5%. Visitor-to-lead rate: 2%.

Starting from the goal, calculate the additional revenue needed,
the corresponding number of customers, leads, and visitors,
then propose the acquisition channels and the resulting budget.
Present the calculations in a table.

Your Turn

Compare two ways of asking. Choose a calculation or decision problem from your week. Pose it a first time with the fastest reasoning level and the formula “reason step by step.” Pose it a second time with a higher reasoning level, all the data, your success criteria, and a verifiable response structure. Compare the accuracy and, above all, the ease with which you can control each response.

Official sources: Reasoning best practices (OpenAI API documentation) · GPT-6 and other models in ChatGPT · Prompt engineering best practices for ChatGPT (OpenAI Help Center) · Prompting (ChatGPT documentation)

Key Takeaways

• ChatGPT models reason internally: “think step by step” is useless and can even be harmful
• When a problem is difficult, raise the reasoning level rather than adding formulas
• Provide all data, the assumptions to keep, and your success criteria
• Structure the expected response so you can verify it: calculations in a table, ranked causes, cited sources
• To plan, start from the goal and work backward to the means