When a thinking model is worth the wait

Most of the models in our catalogue can be told how hard to think before they answer. At our 19 October 2026 check, forty-one of the fifty-nine models exposed that control. Turn it up and the model writes out a working pass first, checking itself, before it says anything to you. That costs you time, sometimes a few seconds, sometimes closer to a minute.

The trade is time against the quality of the answer, and it does not fall the same way every time. For a large share of what people ask in a day, the extra seconds buy nothing you would notice. For a smaller share, they are the difference between an answer that is right and one that is confidently wrong.

Below are eight kinds of task, each with the same four lines and a verdict, including the ones where the verdict is no.

What the wait actually buys

When a model reasons, MultiChats shows you the working pass. While it runs you see a "Thinking..." line above the reply. When it finishes, the working collapses into a block labelled "Reasoning", or "Reasoning · 3 steps" if it took more than one. Open it and read what it did.

That is the mechanism, and it is also the failure mode. A model told to think hard will think hard about anything, including a request that needed no thought at all.

Simon Willison ran Qwen 3.8 27B locally in August 2026 and found its reasoning effort set to the highest level by default, which he called "a hilarious default" and "absolutely not a good way to run the model, especially on consumer hardware". He asked it to draw an SVG of a circle. It deliberated for minutes and returned an "absolutely beautiful" animated circle that was, in his words, "entirely not what I had asked for!"

We do not carry that model. Effort applied to the wrong task makes the answer longer without making it better.

Eight everyday tasks, one verdict each

A multi-step calculation

Looks like: Splitting a bill where two of six people shared a starter. Converting a recipe from cups to grams and then halving it.

What the thinking changes: Arithmetic chains are where a fast answer fails quietly. A model answering immediately commits to the first plausible number and builds on it. A working pass lets it carry the intermediate values and catch a slip before you see the total.

What you will see: Open the Reasoning block and the steps are there in order, with the numbers. If the visible answer disagrees with the model's own working, you have caught it, which is reason enough to open the block on anything you plan to act on.

Verdict: Worth the wait. This is the clearest yes on the list.

A plan with dependencies

Looks like: A move-out checklist where the deposit inspection has to follow the deep clean. A two-week trip where one booking gates another.

What the thinking changes: Ordering constraints are what a fast answer flattens. It will give you a list, the list will look right, and step four will quietly depend on step seven.

What you will see: A thinking pass usually names the constraint out loud in the Reasoning block, something like "the visa has to be applied for before the flight is booked", then builds the order around it.

Verdict: Worth the wait, and the margin grows with each added constraint.

A question with a trap in it

Looks like: Anything phrased to invite a fast wrong answer. Riddles, questions built on a false premise, "which weighs more" comparisons.

What the thinking changes: A model answering quickly matches the shape of the question and returns the answer that shape usually takes. A working pass gives it room to notice the shape is the trap.

What you will see: In the Reasoning block you can watch it reach the obvious answer and then reject it. It is the most legible demonstration of what the setting does.

Verdict: Worth the wait. To watch the control earn its keep in one message, send it one of these.

A rewrite

Looks like: Shortening an email. Changing the tone of a paragraph. Turning bullet notes into prose.

What the thinking changes: Very little. Your text and your instruction are both on the table already, so there is no hidden structure to work out. Deliberation adds passes over a decision that was never hard.

What you will see: In the message stats the reasoning token count climbs while the output reads much as it would have anyway.

Verdict: Not worth the wait.

A factual lookup

Looks like: A date, a capital city, a definition, the spelling of a name.

What the thinking changes: Nothing useful. If the model has the fact, it has it. Thinking longer about a fact it lacks gives you a longer route to the same wrong answer, delivered with more composure.

What you will see: A short Reasoning block that mostly restates your question back to itself. If the fact is one that moves, the control that helps is web search, which is a separate button.

Verdict: Not worth the wait.

A quick summary

Looks like: Compressing an article, a long thread, or a page of notes into a few lines.

What the thinking changes: Not much, because the material is in front of the model. Summarising is a reading task with the source supplied, and fast models are good at it.

What you will see: A longer pause before the first word, for output you would struggle to tell apart. The picture changes the moment you stop asking for a compression and start asking the material a question, as in "which of these arguments is weakest".

Verdict: Not worth the wait for the summary itself, worth it the moment you start interrogating the source.

Code

Looks like: Chasing a bug whose cause you cannot see. Choosing between two ways to structure something. Writing a function you have specified down to the signature.

What the thinking changes: It splits, and this is the one entry whose verdict has two halves. For a bug with a non-obvious cause, or a design call with real tradeoffs, the working pass does the job you wanted done. For code you have already fully specified, it deliberates over a settled decision.

What you will see: On a debugging prompt the Reasoning block reads like an investigation, with hypotheses raised and dropped. On a boilerplate prompt it restates your own prompt in longer words. The block tells you which one you are in.

Verdict: Worth the wait when you are stuck, and skippable when you are dictating.

An ambiguous question

Looks like: "Is this contract fair?" "Should I take the job?" Any question where a competent answer starts by naming what is missing.

What the thinking changes: This is the case most people never think to test. A fast model resolves ambiguity silently: it picks one reading, answers it well, and never mentions that it chose. A working pass is much more likely to surface the fork, either by asking you or by answering both readings.

What you will see: The Reasoning block naming two interpretations. If it names both and answers only one, that is useful too: it tells you which sentence to go back and pin down.

Verdict: Worth the wait, with a limit. It makes the ambiguity visible. Deciding which reading you meant is still your job.

Where the control lives, and who can move it

In the mobile app it is a pill labelled "Reasoning" in the row under the message box, on forty-one of our fifty-nine models. On the web it is a lightning-bolt button in that same row, and it is missing until you turn Intent Detection off. That setting is on for every new account and decides for you when to search, generate an image or adjust reasoning. Its switch is in Settings, under Tool Preferences.

Open the menu and you get Default, where it sits if you have never touched it, then Auto, which hands the depth decision to the model, then Low, Medium and High. "Off" appears on some models and not others, worth knowing before you go hunting for it. Twenty-seven of the forty-one declare that their thinking cannot be switched off, so those menus have no Off row at all. Claude Sonnet 5.5, Claude Opus 5.5, GPT-6.1 Sol, Gemini 3.8 Flash and Grok 4.3 Reasoning are among them. On those, Low is the floor.

The model makers decide that, and we pass it through. Anthropic goes further on its newer models, where thinking is adaptive. Its own documentation says that with adaptive thinking, "Claude decides whether and how much to think on each request, and at lower effort settings it may skip thinking entirely on easy inputs". What you see from the outside is simpler: the level you choose does not always change how long it thinks. Set one, then open the Reasoning block and read what actually happened.

Three OpenAI models push the other way. On GPT-6 Luna and GPT-5.6 Terra, our High sends OpenAI's top effort tier, and on GPT-5.5 it sends the tier below that. If High feels noticeably slower on those three, that is why.

On free accounts the control is visible and locked wherever it appears. You can see the button and its current level, and choosing a different one opens the upgrade prompt instead of the menu. The web tooltip says so directly, reading "Reasoning effort (Default) - Pro". Six of the free models that reason also carry a Medium ceiling enforced on our side, so a request that went around the interface would still not go higher. Reasoning itself still happens on those models, at whatever depth the model maker set. The dial is the Pro part.

Try it on two questions

There is no side-by-side view in the app, so this is a sequential test: send the same prompt twice, once at Low and once at High, and read both.

Take one prompt from the top of the list, a calculation or a plan or a trap question, and one from the bottom, a rewrite or a lookup. On the first you will usually see the difference immediately, often in the Reasoning block before you reach the answer. On the second you usually will not, and learning that is what stops you waiting on answers that were never going to change.

One thing worth settling while you are in there: turning effort up makes a model more careful with the material it already has. It does not hand it material it lacks. Check missing information separately from reasoning effort.