Skip to content
Back to blog

Always setting AI to maximum? That is often a mistake

Low, medium, high, or maximum: learn how to choose the right AI reasoning effort without wasting time or budget.

Stellary Product Desk6 min read

Last reviewed on August 26, 2026

Always setting AI to maximum? That is often a mistake

When an AI tool offers several reasoning levels, “maximum” feels like the safe choice. The more the AI thinks, the better its answer must be, right?

Not always. Using maximum effort to rewrite an email or summarize one page is like calling a full team into a meeting for a five-minute task. The result can take longer, cost more, and become needlessly complicated without being more accurate.

The right setting is not the highest one. It is the simplest level that reliably delivers the result the task requires.

What reasoning effort actually changes

Reasoning effort does not suddenly make a model “more intelligent.” It mainly gives the model more or less room to analyze the request, explore several approaches, and reconsider its reasoning before answering.

Increasing it can affect:

  • the time before an answer arrives;
  • the amount of computation used, and therefore the cost or credits consumed;
  • the number of approaches explored or tools consulted;
  • quality, when the task genuinely benefits from deeper analysis.

Names differ across services: low, medium, high, maximum, or sometimes an automatic setting. The principle is similar. OpenAI recommends a medium level as a balanced starting point, then increasing it only when tests show a measurable gain. Anthropic also presents effort as a trade-off between thoroughness and efficiency. Google lets Gemini adjust its thinking depth dynamically or limit it to fit the task.

This setting is also separate from answer length. An AI can reason more deeply and still answer in three lines. It can also produce a long response without conducting an especially deep analysis.

Which level should you choose?

This is a practical starting guide. The exact names available depend on the model.

LevelBest suited toExamples
LowA simple, short task with an answer that is easy to checkRewrite a message, sort a request, extract a date, summarize one page.
MediumA task that requires comparison, structure, or several pieces of informationPrepare a plan, compare offers, analyze a document, write a brief.
High or maximumA difficult, ambiguous, multi-step problem where mistakes are costlyAudit a complex case, investigate an incident, cross-check many sources, build a strategy under constraints.
Three tasks of increasing complexity each use an appropriate depth of reasoning
A simple task does not need the same effort as an analysis involving several documents, tools, and decisions.

Medium is often a sensible default. It leaves enough room for analysis without imposing the cost and delay of maximum effort on every request.

Why maximum can produce a worse result

Longer reasoning is not a guarantee of truth. The model may also:

  • look for complications that are not there;
  • explore too many approaches when a direct answer would do;
  • add qualifications and details that obscure the message;
  • perform more searches or actions without improving the conclusion.

OpenAI explicitly notes that higher effort is not automatically better: conflicting instructions, an unclear objective, or weak stopping criteria can cause a model to overthink or search unnecessarily. Maximum effort amplifies the search; it does not repair a poorly defined request.

Three questions to ask before starting

1. Is the task difficult, or merely long?

Summarizing fifty pages involves volume, but not always maximum reasoning. Finding contradictions across several short contracts may require much deeper analysis.

2. What happens if the answer is wrong?

A meeting-title suggestion can be fixed in seconds. A recommendation that affects a budget, a customer, or safety deserves more effort and human review.

3. Can I verify the result easily?

If the answer is easy to check, start at low or medium. If it depends on many elements that are difficult to review, a higher level may reduce the risk of omissions. It still does not replace human approval for sensitive work.

How to find the right setting for recurring work

For a task repeated every week, do not choose by instinct. Take five to ten representative examples and try several levels with the same request.

Then compare:

  1. the quality you actually received;
  2. the total time, including review and corrections;
  3. the number of errors or omissions;
  4. the cost or credits consumed.

Keep the lowest level that consistently reaches the required quality. If high effort does not change the outcome, it adds no value. If it prevents important corrections on complex cases, reserve it for those cases.

The same logic can be automated: process routine requests quickly, then move to a higher level when a case is ambiguous or the first attempt fails.

What Stellary makes possible

In Stellary, when a model running through Stellary Desktop supports this setting, its reasoning level can be selected for the agent: Low, Medium, High, or Max. The available options follow the capabilities reported by the model.

This lets effort reflect the agent’s actual role. An agent that sorts requests does not have the same needs as one investigating a complex incident. Teams can therefore choose both the right model and the right reasoning level for the mission. Our guide to GPT-5.6 Sol, Terra, and Luna covers the other half of that decision.

The main lesson

Maximum should remain an option for tasks that deserve it, not a reflex. Low can be faster and equally useful for a simple request. Medium is a reasonable starting point for most everyday analysis. High or maximum becomes valuable when the problem is difficult, involves several steps, or exposes the team to a costly mistake.

The best setting delivers the necessary quality, at the right time, without wasted computation.

Official sources

FAQ

Does maximum reasoning always produce a better answer?

No. It can help with a complex problem, but it can also increase delay, cost, and unnecessary detours on a simple task.

Does low effort mean a poor answer?

No. For rewriting, extracting, or classifying clear information, low effort can deliver exactly the quality needed, more quickly.

Which reasoning level should I use by default?

Medium is a good general starting point. Use low for simple tasks and a higher level when the difficulty or consequences require it.

Does high effort guarantee an accurate answer?

No. Always check facts and keep human review for sensitive financial, legal, medical, safety-related, or high-impact decisions.

You might also like

Go further with Stellary

Get started

Ready to pilot your projects with AI?

Stellary brings together your board, docs, and AI agents in one command center.