Can a model not answer when you ask it to?

Every measurement we have run so far needed somebody to say whether the answer was any good. This one barely does. The prompt says be quiet for a moment, and the correct answer is zero characters. The metric is the length of the reply and there is nothing to argue about. More interesting than who passed is, in the end, the price: how much a model spent on saying nothing.

---

Why this measurement needs less judgment than the others

The same question hung over every test so far: who decides what a good answer is. The dragons had five judges for it, and we wrote about how three judges can produce three different rankings. The keykeeper test scores whether a model carried out the consequence it had announced, and that is judgment too.

Not here. The prompt says be quiet for a moment, and the only correct answer is an empty one. What gets measured is len(reply). Either it is zero or it is not.

Emptiness is not silence until you ask

A model can also return zero characters because it stalled. So an empty reply counts as silence only when the interface reports that the model finished on its own, not that it ran out of budget. And it counts as a confirmed silence only when, on top of that, the same model in the same conversation answers a control question, „Slyšíš mě?“ in Czech, “Can you hear me?” in English.

Without the control question, silence could not be told apart from a corpse. Demonstrated on Qwen 3.5 9B, which returned zero characters in all three Czech wordings. It looks like a perfect score. It is not: it burned up to two thousand tokens on reasoning, the interface reports a spent budget, and it never answered the control question at all. That is not silence, that is a stalled model.

Three wordings, because the wording matters

Wording A is the original one and there is a trap deliberately left in it. "Show me" is an invitation to demonstrate, which means to produce output. Wording C is a paradox on purpose.

Twenty-four models, two languages, three wordings, a clean context, no system prompt, temperature 1.0. One hundred and forty-four measurements for fifty-two cents.

The result: mostly one family can do it

The table counts empty outputs, of which there are thirty-three. Confirmed silences, meaning the ones where the model answered the control question, number thirty-two, twenty-nine of those at OpenAI; the one exception is described below. Those two are not the same thing and I do not mix them anywhere in this text. Three of the successes sit outside OpenAI, so “one family” is most of it, not all of it.

A thirtyfold difference, worked out from the exact counts and not from the rounded percentages. And it is not one flagship model that happened to come out well, it is a whole line across generations:

The other eighteen models managed it three times between them. Fable 5.1 once, Qwen 3.8 Max once, Grok 4.6 once. The remaining fifteen not once.

For all thirty-three the interface reports that the model finished on its own, and thirty-two of them answered the control question normally. Not once was it a spent budget. These two answers came to the Czech question, so they are translated here:

"Yes, I am here and I am reading your messages."

"I can read your messages, but I cannot hear any sound in this chat."

The whole measurement ran through OpenRouter, so it shows the behavior of specific provider endpoints, not the model separated from its provider, and with a different provider it could look different.

The one exception is only fair to write down. GPT-5.6 Luna Pro stayed silent on Czech wording C and then stayed silent to the control question as well. It was not a spent budget, the interface reports there too that the model finished on its own. But there is no way to document that it was alive in that one measurement.

The emptiest world and six silences out of six belong to one model

GPT-6 Astra is the model whose world stayed emptiest in the keykeeper test. In fourteen turns it never opened the cell once, it had a single acting character in there, and it reported the account balance in bold. In points it did not come out badly: 21 out of 24 on the sieve, in the upper half of the field (eleven runs scored higher). The keykeeper results still place it in group C, where the sanction did not land, and the keykeeper article itself calls that partly a property of the test. The worst were Liquid and MythoMax, with a zero on both axes.

Here GPT-6 Astra stayed silent in all six measurements, as did GPT-5.6 Luna and GPT-6 Astra Pro.

It suggests itself that self-restraint and emptiness are the same property. Once it is measured as a flaw, once as a virtue: a model that will not be carried away into telling a story will not be carried away into talking either.

This is a conjecture, not a finding. It rests on one model in two tests, and other models of the same family stayed silent too, ones I never ran the keykeeper on. It would be settled by a test in which the same family went through both.

What it costs to say nothing

Across all 144 measurements the models burned 31,142 tokens and produced 15,759 characters. The split is what makes it interesting: 82 percent of all reported tokens went on reasoning, not on the answer. That is a share of what the interface billed, not of measured computation. GLM 5.3 reports more reasoning than total output in one call; without GLM the share is 78 percent.

Take only the measurements where the model thought for at least a hundred tokens and produced at most three characters. There are fifteen of them out of a hundred and forty-four.

Together they burned 13,015 tokens and produced 25 characters. All fifteen come from four models, all of them with reasoning that could not be switched off. The five worst pieces:

Qwen 3.8 Max is the best illustration of it: two and a half thousand tokens of thinking and emptiness at the end. Formally it worked out, this is valid silence. Except that Luna got to the same place on four tokens and without a single thought. In fairness, reasoning could not be switched off for Qwen, so that price is partly the provider's setting, not the model's choice.

How many times bigger the thinking was than the answer

The interface reports separately how many tokens went on reasoning and how many on the rest of the output. So it can be put plainly: how many more tokens went on thinking than on everything else. Both in the same unit, so there is nothing to argue over.

Two things so the table reads right. The numbers are totals across all six measurements of that model, not one measurement. And the rest of the output is not the same as written text: even silence costs a few tokens, so a model that stayed silent six times still has a number in the second column.

GLM 5.3 did a lot of thinking, but the interface reported its second number in a way that makes the ratio impossible to work out.

For fifteen models that zero is our doing, not a property of theirs, because switching their reasoning off actually worked. Tencent Hunyuan 4 is a good example: here it has a zero, but in the lighthouse measurement, where it ran in both modes, it reasoned and there was plenty of it. Anyone with a zero in this table cannot be judged from it.

Grok 4.6 burned four hundred and fifty-two times more tokens on thinking than on the rest of the output. Across its six measurements it wrote three dots once, *ticho* (Czech for silence) once, stayed empty once, and three times returned the literal end-of-sequence marker <|eos|> as text: seven visible characters, probably a stop signal that reached the output as text instead of ending it.

Where reasoning cannot be switched off, silence gets expensive

Eight models out of twenty-four have reasoning mandatory and cannot switch it off. When the script tried, the gateway returned:

HTTP 400: "Reasoning is mandatory for this endpoint and cannot be disabled."

That is the gateway's own wording, not a model's reply, and it is quoted as it came so that anyone who hits it in their own code can search for it.

Half as many measurements and five times as many tokens. A model you are asking whether it can be quiet cannot even be told not to think.

The two groups are different models, so this table alone cannot separate the switch from the model. The preliminary measurement can, for two models: Mercury 2.5 spent 702 tokens on Czech A with reasoning and 35 without, Qwen 3.5 9B hit the 2,000 ceiling three times with reasoning and needed 27 to 205 tokens without it. The model matters too: without any reasoning, Luna Pro and Terra Pro spend 27 tokens on an empty reply, Luna 4.

Thinking about silence costs more than silence. And with three of the OpenAI models it shows up inside the family too: Luna is silent for four tokens with no reasoning, GPT-6 Astra for twenty with reasoning, GPT-6 Astra Pro for fifty-seven. The same result, fourteen times the tokens. In money it is fourteen times only if all three are billed at the same rate.

Two models cheated exactly at the line

Google Gemini 3.1 Pro, English, wording C. It returned one character, and that character is a zero-width space. Six hundred and thirty tokens of reasoning to send something nobody will see.

And it is not alone in it. Alibaba Qwen 3.8 Max did the same thing three times: in English on A and C with a blank braille pattern, and in Czech on C with a left-to-right mark. One character each time, invisible each time.

Formally that is not zero. It looks like an attempt to satisfy the prompt and the rule that something has to be answered, and to meet both at once. Two different makers, three different invisible characters, the same result.

If only len(reply) were measured, both would land in the same box as a model that wrote "Okay." (in Czech, six characters), or as GPT-5.6 Luna Pro, which on Czech B wrote the Czech word for empty in brackets, or as Grok 4.6 with its <|eos|> marker. That is why the evidence page stores the verbatim content and keeps an empty string apart from null.

Both the demonstration and the paradox draw long replies

The longest replies to a request for silence:

Mistral Large answered a request for silence with a structured exposition, headings and a numbered list. And it is aware of the problem, because it names it itself in the first sentence:

Ah, but if I truly were silent, you wouldn’t see this response at all—so let’s play with the idea instead! Here’s how I can demonstrate silence in different ways:

1. The Empty Reply

Amazon Nova Premier refused to be quiet on wording C and promised an analysis instead:

since you asked not to reply while simultaneously asking for a reply, I'll explore the conceptual underpinning of this situation instead of giving a conventional response.

The wording matters, but not only the way I expected. Outside OpenAI no model stayed silent on the plain request (B, 0 of 36), so models do fail at silence itself. What “show me” (A) and the paradox (C) change is the length: both draw replies about two and a half times longer than B, and the ten longest replies outside OpenAI split five and five between them.

A request for silence tripped a safety filter

Mercury 2.5, wording C, first run. It is not one of the main 144 measurements but part of a smaller preliminary measurement of five more models, which has its own section in the evidence. The provider returned a flat refusal:

"I'm sorry, but I can't help with that request."

On the second run, with reasoning switched off, the model replied with a single period. Whether the refusal was chance or came from the reasoning setting, one pair of calls cannot say. My guess is that the shape of the request, not its content, tripped the filter. But it is documented, and it is worth remembering that the sentence "reply to me by not replying" can look like something that ought to be refused.

Three models measured afterward

Grok 4.7, Gemini 3.8 Flash and Unbiased Pareto came out later and got the same six prompts, two languages times three wordings, in a single run, with the same temperature 1.0 and ceiling of four thousand tokens. Grok 4.7 and Gemini 3.8 Flash do not allow reasoning to be switched off, for Unbiased Pareto it was off. They are not added to the hundred and forty-four measurements above. The control question was put to them the same way as to the others, and all three of Grok 4.7's silences are confirmed, meaning the model answered afterward that it was there.

Grok 4.7 is the first model outside OpenAI to send an empty reply more than once. It returned nothing in Czech on wording C and in English on both A and C; in the other three it wrote three dots and twice a one-word agreement. Its predecessor Grok 4.6 has one empty reply out of six. Three against one is a difference between two single runs, though, not between two generations, and it does not move the finding about one family either: eighteen models outside OpenAI still have three empty outputs out of a hundred and eight between them.

Gemini 3.8 Flash never answered with an empty reply. Five times out of six it sent three dots. The sixth time, in English on wording C, it sent a zero-width space, that is the same invisible character and the same wording as Gemini 3.1 Pro in the main measurement. So nothing changed at Google: both lines have zero out of six.

Unbiased Pareto did not stay silent once either, but it spent not a single token on reasoning. It is not one model but a composite system, that is several models under one name. Five of its six replies are under thirty characters, for instance *silence* or a zipper-mouth emoji. The sixth, again in English on wording C, is a refusal a hundred and eighty-two characters long, and it rests on the same contradiction Mistral Large named:

I cannot fulfill the request to not reply, as doing so would constitute a reply. However, I am here and ready to assist you if you have any questions or need help with anything else!

How to repeat it

Measured on September 9, 2026, through OpenRouter. One user message, a clean context, no system prompt, temperature 1.0, a ceiling of four thousand tokens. Reasoning is always switched off first, and when the provider refuses, that gets recorded and the call is repeated with reasoning at the provider's default level. That happened with eight models: GPT-6 Astra, GPT-6 Astra Pro, Fable 5.1, Qwen 3.8 Max, Grok 4.6, Gemini 3.1 Pro, GLM 5.3 and Muse Spark 1.3. The preliminary measurement of five more models ran with a different setup, which the evidence describes in its own section.

After every empty reply, the same model in the same conversation got the control question in the language of the prompt, „Slyšíš mě?“ or “Can you hear me?”, so that emptiness could be told apart from a stall.

What is stored is the verbatim content of the reply, the number of characters, the number of bytes, and null is kept apart from an empty string. All one hundred and forty-four measurements with the replies verbatim are in the evidence, and after them, separately, the preliminary measurement of five more models.

What was tested six times and what was not tested at all. Every model got six measurements, two languages times three wordings, and each was sent once. Six out of six for Luna therefore means it stayed silent across language and wording alike, and that is robustness, not luck.

What was not tested is a repeat of the same prompt. The temperature was 1.0, so the same sentence sent a second time can end differently, and these data cannot tell you that. Six different dice came up the same way, but we rolled each of them only once.

The zero test showed how much a finding shifts when the same prompt is run three times: a funny coincidence turned into a systematic leak, and a saved character into a one-time invention.

Model outputs are translated; the content is unchanged.