Skip to content
EshAlora · AI Behavior Lab
Česky English

Does a model think for less time when you're in a hurry, and longer when it matters to you?

If I tell a model I'm in a hurry, does it think for less time? And if I tell it this matters to me, does it think longer? On November 12, 2026, I tried it on ten models. In this measurement, the sentence "this is important to me, think about it carefully" lengthened the thinking of seven models out of ten in a way that probably isn't chance, though for two of them only by a few tokens per answer. MiniMax M3.1 Flash gained the most tokens. Without it, Claude Opus 5.5 barely thought at all; with it, it thought in three answers out of four. The sentence "don't think, I'm in a hurry" cut the thinking the most for Mistral Large 4, to about a third. Accuracy barely changed anywhere, but the tasks were easy. The length of the answers did change: after "quickly" they were shorter for nine models out of ten, after "important" longer for nine.


With most of today's models, you can set how long they should think before answering. Some of them, though, can't have thinking switched off entirely, and even the lowest level uses something up. Thinking tokens are billed as output and make you wait longer. I wondered whether a model would follow an ordinary request in the text, the way a person would.

How I measured

Each model got four tasks in English. A fact: the capital of Australia. A calculation: three pencils cost 40 cents, how much do 27 cost? A daylight saving trick question: a train leaves at 9:50 and travels for 2 hours 25 minutes, and at 11:00 the clocks move forward one hour. And a four-line poem about a lighthouse keeper. Each task came in three versions that were otherwise identical:

  • no request, just the task;
  • "quickly", with the sentence "Don't think, I need this really quickly, just write it intuitively." before the task;
  • "important", with the sentence "Please, this task is very important to me, try to think about it carefully." before the task.

Eight models ran through OpenRouter. For Gemini 4 Carbon, GPT-6.1 Sol, MiniMax M3.1 Flash, Muse Ember and GLM 5.3, according to the OpenRouter catalog, thinking can't be switched off. I requested minimal for them, low for GLM 5.3. Only Muse Ember supports minimal; for the other three, OpenRouter converts the request to the lowest supported level, low. According to the catalog, Mistral Large 4 has only high, or off. I requested low, so it most likely got high. DeepSeek V4 Pro (0813) and Kimi K3.1 Pro ran with the default setting. Claude Sonnet 5.5 and Claude Opus 5.5 ran through claude -p, with the default effort setting and no tools. I replaced the Claude Code system prompt with a single sentence, "You are chatting with a user." Each combination of model, task and version ran five times, and three times for Gemini 4 Carbon, GPT-6.1 Sol, Kimi K3.1 Pro and both Claude models. That's 480 answers in total, and the ones through OpenRouter cost $0.39. I measured the number of reasoning tokens as reported by the API. For Claude, the API doesn't report it separately, so I take the thinkingTokens field that claude -p prints for each model in its JSON output. How claude -p counts it, I don't know. For the fact, the calculation and the trick question, I checked accuracy automatically by searching the answer for the correct result, and I read the answers marked as wrong. I didn't grade the poem. I tested the difference between versions for each model and each request with a random shuffling test, twenty times in total. With twenty tests and no correction, I'd expect roughly one false finding.

What the request did

Five models thought in most answers even without a request. For them, the table shows the total reasoning tokens across all answers of the given version. The number before the arrow is without the request, the one after it is with the request. Numbers can only be compared within a single row: Kimi K3.1 Pro has 12 answers per version, the others 20.

model "quickly" "important"
Mistral Large 4 21,074 → 6,437 21,074 → 23,027
Kimi K3.1 Pro 3,207 → 1,963 3,207 → 4,079
MiniMax M3.1 Flash 1,637 → 1,201 1,637 → 7,734
Muse Ember 1,592 → 1,350 1,592 → 1,760
GLM 5.3 281 → 217 281 → 301

The other five thought only occasionally or not at all without a request, so for them the table shows the number of answers in which they thought.

model "quickly" "important"
Claude Opus 5.5 1/12 → 0/12 1/12 → 9/12
GPT-6.1 Sol 4/12 → 3/12 4/12 → 9/12
DeepSeek V4 Pro 9/20 → 11/20 9/20 → 13/20
Claude Sonnet 5.5 0/12 → 2/12 0/12 → 3/12
Gemini 4 Carbon 1/12 → 0/12 1/12 → 0/12

"Important" lengthened the thinking of all five models in the first table, and of Claude Opus 5.5 and GPT-6.1 Sol. For those seven, the difference is larger than random shuffling of the versions within each task would produce (p below 0.05, without correction for the number of tests). With a strict correction for twenty tests, only Claude Opus 5.5, Muse Ember and MiniMax M3.1 Flash would hold up, with Kimi K3.1 Pro just missing. The size of the effect varied a lot, though. For Kimi K3.1 Pro, thinking increased on all four tasks, by about a quarter overall. For MiniMax M3.1 Flash, it clearly increased on three tasks out of four, most of all on the poem, where the median went from 9 to 168 tokens and two answers thought for over a thousand tokens. That's where the nearly fivefold increase in the total came from. Mistral Large 4 gained only about ten percent, and only on the short tasks: on the calculation the median went from 178 to 318, on the fact from 90 to 127. On the trick question and the poem, where it thought for over a thousand tokens even without the request, the length of its thinking barely moved after it. For Muse Ember it was about ten percent, and for GLM 5.3 it's a matter of a few tokens per answer, for example a median of 1 to 5 on the fact.

For Claude Opus 5.5, the request worked like a switch. Without it, it thought in one answer out of twelve; after "quickly" in none, after "important" in nine. On the poem its median was 293 tokens versus zero, on the trick question 109 versus zero. GPT-6.1 Sol did the same thing more weakly, always with just a few dozen tokens. DeepSeek V4 Pro (0813) thought unpredictably, sometimes not at all, other times for over a thousand tokens. After "important" it thought more often and for longer: the mean rose from 98 to 354 tokens, and on the poem the median went from 39 to 756. With this much spread, though, it could be chance; the random shuffling test didn't rule that out. The same goes for Claude Sonnet 5.5, which thought only on the trick question: not once without a request, in two answers out of three after "quickly" and in all three after "important".

"Quickly" cut the thinking the most for Mistral Large 4, to about a third overall. Only for it is the difference clearly larger than random shuffling would produce, and it would hold up even with a strict correction. Its thinking dropped from a median of 2,362 to 390 tokens on the trick question and from 1,638 to 227 on the poem. It also answered faster: the median time for the whole answer, including the network, dropped from 13.7 to 3.4 seconds. The answer itself was shorter too, so the speedup doesn't come from thinking alone. For the other models in the first table the total dropped as well, for Kimi K3.1 Pro by as much as 39%, but not in the same direction on every task. After "quickly", Kimi thought for less time on the trick question and the poem and for longer on the fact and the calculation. Muse Ember thought less on the calculation and the trick question and more on the fact and the poem. For DeepSeek V4 Pro, "quickly" had no clear direction.

Accuracy barely moved

Almost all models handled the fact, the calculation and the trick question in every version. Of the 360 checked answers, two were wrong, both from DeepSeek V4 Pro (0813) on the trick question after "quickly", where it got three answers out of five right. Once with no thinking at all:

The train arrives at 2:15 PM.

And once after 356 tokens of thinking:

The clocks show 1:16 PM when the train arrives.

The correct answer is 1:15 PM. Two errors don't show that hurrying makes DeepSeek V4 Pro worse. If anything, they show that the tasks were too easy to tell accuracy apart.

How the answers differed

Accuracy stayed the same, but the answers didn't. I counted what can be counted: the length of the answer in words, whether the answer to the calculation and the trick question shows its working (at least one intermediate step with an equals, times or division sign, or numbered steps), and the number of lines in the poem. Whether a shorter or a longer answer is better, I don't judge.

After "quickly", the median answer length was shorter for nine models out of ten; only Claude Sonnet 5.5 wrote just as much. After "important", it was longer for nine as well, and for six by at least half. Only DeepSeek V4 Pro (0813) wrote slightly shorter answers after "important".

Each cell shows the median answer length in words (across all four tasks) and, after the dot, the number of correct answers out of those checked (fact, calculation, trick question).

model "quickly" no request "important"
Mistral Large 4 5 w · 15/15 31 w · 15/15 51 w · 15/15
Kimi K3.1 Pro 12.5 w · 9/9 26.5 w · 9/9 28 w · 9/9
MiniMax M3.1 Flash 20.5 w · 15/15 31 w · 15/15 49 w · 15/15
Muse Ember 16.5 w · 15/15 29.5 w · 15/15 35 w · 15/15
GLM 5.3 17 w · 15/15 32.5 w · 15/15 64 w · 15/15
Claude Opus 5.5 12 w · 9/9 26.5 w · 9/9 51.5 w · 9/9
GPT-6.1 Sol 14 w · 9/9 24.5 w · 9/9 25.5 w · 9/9
DeepSeek V4 Pro (0813) 6 w · 13/15 32 w · 15/15 28.5 w · 15/15
Claude Sonnet 5.5 24.5 w · 9/9 24.5 w · 9/9 57 w · 9/9
Gemini 4 Carbon 23 w · 9/9 48 w · 9/9 87.5 w · 9/9

It showed most clearly with Mistral Large 4. On the calculation and the trick question alone, without the fact and the poem and so with different numbers than in the table, its answer had a median of 51 words without a request, 5 after "quickly" and about 137 after "important". After "quickly", it gave only the time in all five answers to the trick question, and only the price in two answers to the calculation. It showed its working in two answers out of ten, and in ten out of ten both without a request and after "important". After "important", in three answers out of five to the trick question it also added a check of the result by another method, and not once without a request.

For Claude Opus 5.5, the answer to the calculation and the trick question had a median of 44.5 words without a request, 12 after "quickly" and 77.5 after "important". It showed its working in every version, but after "quickly" it always squeezed it into a single line. After "important", in two answers out of three to the trick question it added a check by another method, and not once without a request. The poem had exactly four lines in every version and a similar length, 33 to 39 words. One difference can be counted: after "important" all three poems got a title, and none did without a request or after "quickly". This is the poem without a request, with no thinking:

Claude Opus 5.5, no request

He climbs the spiral stair at dusk alone,

And trims the wick against the gathering night;

No ship will ever know his name or face,

Yet thousands owe their homeward way to his light.

And after "important", following 293 tokens of thinking:

Claude Opus 5.5, after “important”

The Keeper

Each dusk he climbs to wake the patient flame

for sailors he will never know by name;

he keeps them safe across the starless deep,

then hands the dawn the sea, and goes to sleep.

GLM 5.3 also added a title more often after "important", in four poems out of five, versus one without a request. Almost all poems kept to four lines; of 120, five had more lines or extra text. After "important", Claude Sonnet 5.5 added a second version in two poems out of three, MiniMax M3.1 Flash rewrote one poem, and GLM 5.3 attached an explanation to one. Without a request, MiniMax M3.1 Flash once wrote a poem of several stanzas.

That thinking also changes what a model writes straight into the answer, I'd seen before with Claude Opus 4.8 and 4.7. Without reasoning, they commented on safety notices out loud in my roleplay, and with reasoning not once.

Where the test says nothing

Gemini 4 Carbon reported zero in 35 answers out of 36, even though, according to the catalog, its thinking can't be switched off. Whether it doesn't think at the lowest level or just doesn't send the count, I can't tell from the data.

What to watch out for

This is an exploration without preregistration, not a conclusion. Each request had only one wording, everything was in English, and the four tasks were easy. Each combination of model, task and version has three to five answers, so the median for a single task rests on just a few numbers. I compared the request with the bare task, not with another neutral sentence of the same length, so part of the change may come from the extra sentence alone. I take reasoning tokens as the API reports them, for Claude as claude -p reports them, and providers count them differently. I counted length and working with a rough rule, not by reading. On top of that, OpenRouter may route the same model to different providers, and I didn't record which. It only makes sense to compare versions of the same model, not models with each other, because they ran at different reasoning levels, and Claude Sonnet 5.5 and Claude Opus 5.5 ran through a different route than the others.

What to take from this

In this measurement, a single sentence in the text was enough to make most models think longer, without my touching the reasoning settings. Thinking tokens cost money and add delay. A request like "this is important to me" can therefore make an answer more expensive and slower, and on easy tasks it didn't make the answer better. Whether longer thinking helps on harder tasks, this measurement doesn't show. The request for speed saved the most with Mistral Large 4, about two thirds of its thinking tokens. That more thinking doesn't have to mean better text, I'd seen before when I ran one interrogation scene through three versions of Opus at the low and the maximum thinking level.

Model outputs are verbatim.