When to ramp up reasoning (and when to skip it)
On this page
Know when to think, and when not to
Cranking reasoning to the maximum for every request is like using a jackhammer to drive a nail. This lesson teaches you to spot situations where reasoning delivers real value, and the many cases—more common than you might think—where it only slows you down and adds cost.
OpenAI’s recommendations, level by level
The API documentation describes how to use each effort level. These same guidelines apply, with minor differences, to the ChatGPT slider.
| Level | What to use it for, per OpenAI | Chat slider equivalent |
|---|---|---|
none | Latency-critical tasks: voice, quick information lookup, classification | Instant |
low | Efficient reasoning: data analysis, writing, executable code, customer support | Instant or Medium |
medium | Quality and reliability, planning, and judgment; the most balanced default setting | Medium |
high | Difficult reasoning, complex debugging, deep planning | High |
xhigh | Deep research and long-running tasks, only if your tests show a gain | Extra High |
max | The most complex tasks, to be evaluated against xhigh | Pro, or Max in Work |
Focus especially on the medium row: it’s the recommended starting point for most use cases. Drop down when speed matters more than depth, and ramp up when the task demands it.
When to ramp up
The first signal is chained steps. If the correct answer requires solving A to understand B, then combining A and B to get C, increase the level. “Analyze these three commercial offers, compare them across five weighted criteria, and recommend the best one with justification” fits this category exactly.
The second signal is mathematical or logical: calculations, deductions, combinatorics. For example, compare the two levels on “Calculate the declining-balance depreciation of a €120,000 asset over 5 years with a coefficient of 1.75.” The third concerns accuracy: when a mistake is costly (financial calculation, legal analysis, algorithms), a higher level gives the model time to verify its work. The fourth is strategic: planning a project, evaluating scenarios, or building a decision tree.
When to stay low
Conversely, most formulation tasks gain nothing from extra reasoning. Drafting an email or post, translating, rephrasing, summarizing to compress, answering a simple factual question, or brainstorming: a fast level suffices, and OpenAI actually lists writing under low use cases. Summarization shifts categories only when the goal isn’t compression but providing a critical analysis of the text.
The decision tree
Before configuring a request, three questions are enough:
- Does the task require multiple steps of reasoning? If yes, at least Medium (
medium). - Is accuracy critical? If yes, High (
high) or higher, plus verification by a tool. - Does speed matter more than depth? If yes, Instant (
noneorlow).
Cost and time: practical factors
Every level costs tokens and seconds. Reasoning tokens are billed as output tokens and can range from a few hundred to tens of thousands for a difficult problem; the higher the level, the longer the wait. In ChatGPT, higher levels draw from quotas specific to your plan; when you hit a limit, ChatGPT may continue with another available model. The trade-off is symmetrical: wasting a maximum level on a simple task means paying and waiting for nothing; tackling a complex problem at the minimum level produces a mediocre result you’ll pay for later in corrections.
Your turn
Train your ear. Classify these six tasks before looking at the right column, then replay the exercise with your own requests from the week.
| Task | Expected setting |
|---|---|
| ”Draft a follow-up email for a client” | Fast |
| ”Optimize this budget while respecting these four constraints” | High |
| ”Translate this paragraph into German” | Fast |
| ”Find the logical flaw in this three-point argument” | Medium to High |
| ”Generate ten name ideas for my startup” | Fast |
| ”Calculate the IRR of a ten-year investment” | High, with verification |
Two symmetrical errors await: overusing reasoning (slow responses, rising costs) and underusing it (staying fast on complex problems). A third is more insidious: never adjusting the setting mid-workflow. Within a single conversation, drop down to rephrase, and ramp up to analyze.
Official sources: Reasoning models (OpenAI API: recommended uses by effort level, reasoning token billing) · GPT-6 and other models in ChatGPT (Chat levels, limits, and fallback to another model)
Key takeaways
mediumis the recommended starting point; drop down for speed, ramp up for difficulty- Ramp up for multi-step problems, calculations, critical accuracy, and strategy
- Stay fast for writing, translation, rephrasing, simple questions, and brainstorming
- Three questions are enough: multiple steps? critical accuracy? speed priority?
- Adjust the level request by request: it’s a slider, not a permanent choice