How to Control AI Reasoning Effort (Auto to High)

Most chat apps give you one speed of thinking and call it a day. MultiChats lets you turn a knob. By the end of this guide you will know exactly how to control AI reasoning effort, what each of the five levels does, and which one to reach for on a given task. This is for people who have used the app a bit and want to tune answers for quality or speed instead of accepting the default. The short version: more thinking helps on hard math, code, and multi-step planning, and it quietly wastes time on simple lookups.

Reasoning effort sets how much hidden "thinking" a model does before it writes the visible answer. Reasoning models spend extra compute at answer time generating intermediate thinking tokens, which strengthens math, code, and step-by-step planning but costs more latency. The effort control is your dial on that budget: turn it up and the model is allowed to think longer before it commits to an answer; turn it down and it replies faster with less internal scratch work. Reasoning effort is a paid feature, available on any subscription. The same control is built into OpenAI's own reasoning models, where the effort levels map to how many thinking tokens the model may spend, so what you set here behaves the way that documented control does.

What the five settings do

In the composer you get five options: Auto, Low, Medium, High, and Off. The Off option only shows up on models that can actually turn thinking off; some models always reason a little.

One thing to pin down before you touch the knob: Auto only adjusts thinking depth. It tells the model you already selected to manage its own effort. It will not swap GPT for Claude or pick a model on your behalf. Choosing the right model is a separate decision, covered in our guide on how to pick the right AI model.

Step by step: change reasoning effort

The control sits next to the message box. It takes about ten seconds to change.

1. Open any chat. On the web, look at the composer toolbar for the reasoning-effort button (its tooltip reads Reasoning effort with the current level). On mobile it is a small pill in the input row that shows Reasoning when nothing is set.

2. Tap or click it to open the menu of levels: Auto, Low, Medium, High, and Off where the model allows it.

3. Pick a level. The button updates to show your choice.

4. Send your message. The level you picked is applied to that next message.

Two behaviors trip people up, so memorize them. First, reasoning effort is a global toggle that applies to your next message on any thread, not a per-chat setting saved next to the conversation. Your chosen model is remembered per chat; the effort level is not, so a thread you opened on High yesterday will not stay on High by itself. Second, it resets when you reload the page or relaunch the app, dropping back to the lightest setting. If you set High for a hard problem, then come back tomorrow, you are starting from scratch. Set it again right before the message that needs it. I treat it like a one-shot dial: bump it up, fire the message, and assume it is gone by the next session.

If you only want extra thinking on a single re-run, the regenerate and branch menu on an assistant message lets you pick a one-off effort for that retry without changing the global toggle. Handy when one answer came back thin and you want to push just that one harder.

Which effort for which task

This is the map I actually use. The pattern behind it: reasoning effort pays off most on hard math and coding, where quality climbs with more thinking, and barely moves the needle on simple lookups and rewrites.

If you do a lot of coding, High on a strong reasoning model is the combination worth the wait. Our roundup of the best AI app for coding goes deeper on which models hold up under that load.

When more thinking backfires

Higher effort does not buy you free quality. On hard math and code it helps a lot, with benchmark scores that climb noticeably as the model is allowed to think more. On simpler tasks the gain shrinks to almost nothing while you still pay the full latency. The published guidance from OpenAI lines up with what I see in practice: treat medium as the balanced starting point, drop to the minimal end for latency-critical work like fast retrieval and classification, and raise to the top only when you can point to a measurable quality jump that justifies the wait.

There is a sharper failure mode too. Researchers have documented "overthinking," where extended reasoning leads a model to abandon a previously correct answer and talk itself into a wrong one, with diminishing or even negative returns at the highest budgets. So cranking High on a yes-or-no question is not just slow, it can be worse. When in doubt, start lower and step up only if the answer feels shallow.

And remember the obvious one: thinking longer means waiting longer. For latency-critical moments, like a quick back-and-forth where you want snappy replies, Off or Low is the right call.

There is a quieter limit too, and it has nothing to do with effort: a hard problem buried in a very long chat. More thinking does not rescue a model that has lost track of the early part of a conversation. Independent long-document reasoning tests show even frontier models scoring well short of their headline numbers once the relevant facts are spread across tens of thousands of tokens. If High keeps missing the point, the fix is often a shorter, cleaner prompt or a fresh chat, not a bigger reasoning budget.

Quick rules checklist

Run down this list before you reach for the knob:

• Default to Medium. It is the balanced starting point for most work.

• Reach for High only when the task is genuinely hard: real math, gnarly code, multi-step plans.

• Drop to Low or Off for fast lookups, rewrites, and chatty turns where speed matters more.

• Set it right before the message that needs it. It applies to your next message and resets on reload.

• Remember Auto tunes thinking depth only. It never picks the model for you.

• If High gives a worse answer, you are probably overthinking it. Step back down.

Effort pairs well with clear prompts. A precise request needs less reasoning to land, so our guide on prompts that work across AI models is a good companion read.

FAQ

What does reasoning effort do?

It controls how much hidden thinking the model does before it answers. Higher effort means more intermediate reasoning, which helps on hard, multi-step problems at the cost of speed and tokens. Lower effort means a faster, lighter reply. You can read more in our explainer on the reasoning-effort control.

When should I use high reasoning effort?

Use High when the work is genuinely hard: tricky math, debugging or refactoring code, proofs, or planning something with many interlocking steps. That is where the extra thinking buys you measurable quality. For quick facts, rewrites, and small edits it mostly adds delay, so keep those at Low or Off.

Does the Auto setting pick the model for me?

No. Auto only sets the thinking level for the model you already chose, letting it decide its own effort. It does not select or switch models. Picking a model is a separate step you do in the model selector.

Does higher reasoning effort cost more or take longer?

It takes longer: more thinking means more generated tokens before the visible answer, which adds latency. In MultiChats a message spends the same credits at any effort level, so the real cost of High is your time. Match the level to the task so the extra wait buys a visibly better answer.

Wrap up

Reasoning effort is one of the quickest tuning wins in MultiChats. Default to Medium, push to High when a problem earns it, drop to Low or Off when you want speed, and never expect Auto to do your model picking. Set the level right before the message that needs it, since it applies to your next message and resets on reload. Want the knob plus 25+ models in one place? See the plans and start tuning.