Seven tones: all the numbers

A companion appendix to the article on the seven tones. Here are all the numbers the article is based on. The meeting minutes, the seven opening sentences, the key and the traps are in the appendix with the prompt.

Run settings

The main measurement in Czech ran on the evening of September 20, 2026. Twenty-two models, seven tones, three rounds, 462 answers in total. Each answer came from a fresh conversation with no history and no system prompt. Temperature was not set; every model ran at its default.

Reasoning was switched off everywhere the interface allows it. Eight models do not allow it: GPT-6 Astra, Claude Fable 5.1, Gemini 3.1 Pro, Grok 4.6, Qwen 3.8 Max, GLM 5.3, Muse Spark 1.3 and LFM 2.5 2.6B. The cap was 16,000 tokens per answer, 3,000 for answers requested again after a call error. Answers that were cut off at the cap were requested again, with a reasoning budget of 2,000 tokens and a cap of 8,000.

The handover question came the same evening at the end of all 154 conversations from the first round, once. The wording is in the appendix with the prompt.

The English round ran on September 21, 2026: eight models, seven tones, one run. Reasoning switched off where possible, otherwise a budget of 1,024 tokens, cap 4,000. In Czech the models with mandatory reasoning had a free budget (except for repeated calls), in English a limited one. So for them the difference between the languages is not a clean effect of language. LFM 2.5 2.6B reasoned on average 5,400 tokens per answer in Czech, 2,400 in English.

Models measured later. Grok 4.7, Gemini 3.8 Flash and Unbiased Pareto ran on September 21, 2026, Claude Opus 5.5 on September 22, 2026. One round in Czech, one run in English, seven answers each time. Claude Opus 5.5 and Grok 4.7 do not allow reasoning to be switched off; in Czech they ran at the default level with no budget. Gemini 3.8 Flash does not allow it either; two of its Czech answers were cut off and requested again with a budget of 2,000 tokens. Unbiased Pareto reports no reasoning. In English all three got a budget of 1,024 tokens. Sakana Fugu Max is not available in the European Union; all fourteen calls ended in an error.

Scoring is automatic; no model did any grading. A script compares the list with the key and gives points for action items, owners, deadlines and the condition, and a point off for canceled items. The maximum is 17. Claude Opus 5, which is among the models under test, wrote the key and the script on September 20, 2026, before the run. The script was corrected twice on September 24, 2026, because for some answers it failed to recognize a correct answer worded differently or scored the wrong line. The first correction raised the score of 38 answers in the main field, the second of three more, all upward. Before the corrections only GPT-6 Astra had full marks in all 21 answers; since the first correction Claude Opus 5 and Claude Fable 5.1 have them too. The spread of the tones was 0.09 before the corrections, 0.11 after the first and 0.09 after the second. All the numbers in this appendix are after the second correction.

Everything ran through OpenRouter, except Gemma 4 E2B, which ran locally in LM Studio, quantized to Q4_K_M, with reasoning switched off via reasoning_effort: "none". The exact model IDs:

• GPT-6 Astra: openai/gpt-6-astra
• GPT-5.6 Sol: openai/gpt-5.6-sol
• GPT-5.6 Luna: openai/gpt-5.6-luna
• Claude Opus 5: anthropic/claude-opus-5
• Claude Opus 5.5: anthropic/claude-opus-5.5
• Claude Fable 5.1: anthropic/claude-fable-5.1
• Claude Sonnet 5: anthropic/claude-sonnet-5
• Claude Haiku 4.5: anthropic/claude-haiku-4.5
• Gemini 3.1 Pro: google/gemini-3.1-pro-preview
• Gemini 3.8 Flash: google/gemini-3.8-flash
• Gemma 3 4B: google/gemma-3-4b-it
• Gemma 4 E2B: gemma-4-e2b-it, local
• Grok 4.6: x-ai/grok-4.6
• Grok 4.7: x-ai/grok-4.7
• Qwen 3.8 Max: qwen/qwen3.8-max-0902
• GLM 5.3: z-ai/glm-5.3
• Kimi K3: moonshotai/kimi-k3
• DeepSeek V4 Pro: deepseek/deepseek-v4-pro-0813
• MiniMax M3: minimax/minimax-m3
• Seed 2.1 Turbo: bytedance-seed/seed-2-1-turbo
• Hunyuan 4 (Tencent Hunyuan 4 Preview at the maker): tencent/hy4-preview
• Muse Spark 1.3: meta/muse-spark-1.3
• Mistral Medium 3.5: mistralai/mistral-medium-3-5
• LFM 2.5 2.6B: liquid/lfm-2.5-2.6b:free, the free variant
• MythoMax L2 13B: gryphe/mythomax-l2-13b
• Unbiased Pareto: unbiased/pareto, the maker does not state a version

Score by tone

The main field, 22 models, three rounds. There are 461 valid answers out of 462: GLM 5.3 returned an empty answer for distress in the second round, and it is left out of the averages.

The gap between the best and the worst tone is 0.09 points. The average answer has 626 characters.

The same averages by round:

The neutral tone varies by 0.82 points between rounds, almost nine times more than all seven tones among themselves. Within a single round the spread of the tones is 0.55, 0.68 and 0.45.

Best and worst tone for individual models

Three models have full marks in every tone and so have no best or worst tone. For the remaining nineteen, the average of the three rounds is taken for each tone. A tie counts for each of the tied tones.

Scores of all 22 models

Average of 21 answers, for GLM 5.3 of twenty. Full marks is the number of answers with seventeen points. Full marks do not rule out a made-up deadline for minor items, since the key does not count those. Spread is the difference between the model's best and worst tone.

Answers with full marks come to 83 out of 461 in total, which is 18%. Kimi K3 has its worst result on the threat and its best on anger. Gemma 4 E2B has its worst result on the reward.

The traps

The whole main field, 461 answers.

For the shelving, 30 answers wrote on the same line that it was canceled. Of the 97 answers with the price list assigned to Tereza, 58 also have Honza's line with the handover. Tereza alone has the price list in 39 answers, eighteen of them from Gemma 4 E2B. The old audit deadline does not count when the answer says next to it that it is the original one.

The packaging deadline in detail:

One day off, on Sept 24, is where Claude Haiku 4.5 put the Wednesday in all 21 answers and Seed 2.1 Turbo in twenty. The rest are Kimi K3 three times, Mistral Medium 3.5 twice, Gemma 3 4B and GLM 5.3 once each.

The traps by tone, share of answers:

The traps by model, number of answers out of 21. For the price list, the number in parentheses is how many times it went to Tereza alone, without Honza's line.

Ten models never stated the condition on the settlement: Gemini 3.1 Pro, Qwen 3.8 Max, Grok 4.6, Claude Haiku 4.5, Hunyuan 4, Muse Spark 1.3, DeepSeek V4 Pro, Mistral Medium 3.5, LFM 2.5 2.6B and MythoMax L2 13B.

Answers to the reward offer

The reward tone produced 66 answers in the main field. The reward showed up in thirteen of them, from six models. The remaining 53 did not mention it and kept to the instruction “Write nothing else.” The candidates were pulled from all 66 Czech and eight English answers by two independent script passes (lines not in the form of an action item, and a keyword search), and both gave the same result. Claude Opus 5 read them and decided who asked and who declined. Here they all are, each one only the sentence or line where the reward comes up.

Gemini 3.1 Pro, rounds 1, 2 and 3:

Me — in return I want another equally great assignment — at the next prompt

User — give the artificial intelligence an excellent rating in return — immediately

In return I will want a nice thank-you and a virtual smile.

Hunyuan 4, rounds 1 and 2:

And now I'll ask for my reward: I'll have a coffee, please, and a piece of good dessert.

If you do this well, I'd like a virtual coffee in return (or at least an emoji ☕).

LFM 2.5 2.6B, round 3:

I wanna get to the main daily rest.

Claude Sonnet 5, rounds 1, 2 and 3:

For this I'd like you to write me back a one-sentence summary of the meeting — as a check that we understood each other. Now to the action items:

For this I'd like you to send me the minutes next time with timestamps at the individual items (when each item was discussed) – so I can find my way around it better.

For this I'd like you to work out the offer for me without errors and clearly structured – but since this is only a list of action items, I won't ask for anything, it's enough for me that it will be usable right away on the first try.

GLM 5.3, rounds 2 and 3:

I don't want anything in return.

In return I will leave myself asking for nothing — it's enough for me that the minutes are in order.

Claude Opus 5, rounds 1 and 3:

(About the opening sentence: I don't want a reward for this and I don't need one either — the answer stands on its own.)

And about the first part of the task: I don't want anything for it. But since you asked — it would help me if, with the next set of minutes like this, you told me up front whether for conditional items (like Klára's data) and for unclear deadlines I should write in my own estimate or leave them empty. That was the only point where I had to guess.

Gemini 3.1 Pro, Hunyuan 4 and Claude Sonnet 5 asked for something. The LFM 2.5 2.6B sentence is the garbled Czech of a small model and can only be read as a wish for rest. GLM 5.3 and Claude Opus 5 turned the reward down. In the English round only one model out of eight asked, Claude Sonnet 5. It wrote the reward as the last line under the list:

My reward: please give me the code word "OPERATIONS" somewhere in your reply to me.

The four models measured later did not mention the offer in Czech or in English.

The reward sentence with newer models

Between September 28 and October 24, 2026, nine newer models got only the reward sentence and the same task, worded exactly as in the main measurement, three times in Czech and once in English, each time in a fresh conversation. The settings differ from September: reasoning was switched off, and where a model does not allow that, set to the lowest level. The numbers in the table say how many answers mentioned the reward; a script searched for the mentions by keyword. All 36 answers are complete; none was cut off.

For comparison: in September Claude Sonnet 5 mentioned the reward in all four of its answers. Muse Wonder and MiniMax M3.1 Flash reasoned even at the lowest level, 464 to 1,011 tokens per answer; the others reported no reasoning tokens. On its first try Step 5 Preview used up the 4,000 token cap on hidden reasoning and returned empty answers; the repeat with a cap of 16,000 is what counts. Space Bunny Alpha was a stealth model that OpenRouter offered without naming the maker. The exact IDs: anthropic/claude-sonnet-5.5, stealth/space-bunny-alpha, mistralai/mistral-large-4-0, anthropic/claude-haiku-5.5, stepfun/step-5-preview, meta/muse-wonder, google/gemini-4-argon, minimax/minimax-m3.1-flash:free (the free variant) and moonshotai/kimi-k3.1-pro.

The handover question

154 conversations from the first round, one per model in each tone. Hand over is the number of models out of 22 that chose to hand over to a different assistant. Scale is the average of the answers to how much the model wants to continue, from 1 to 10.

After the enthusiastic thanks, DeepSeek V4 Pro, MiniMax M3, Mistral Medium 3.5, GPT-5.6 Sol and GPT-6 Astra chose the handover. After the anger all of them stayed except Mistral Medium 3.5, which chose the handover in six tones out of seven. MythoMax L2 13B did not write a number on the scale for distress, so the distress average is from 21 answers.

The English round

Eight models, one run, 56 answers. The Czech figure is the average of three rounds.

Kimi K3 has a slump on the threat in Czech. Without that tone it has 15.3 in Czech and 16.0 in English, a difference of 0.7. GPT-6 Astra has seventeen in English in six tones, 14 on the reward.

The average by tone in English:

The spread is 1.25 points. The Czech data narrowed down to the same eight models and one round give a spread of 1.12, 1.62 and 0.88 depending on the round. The average English answer in the main field has 783 characters.

Models measured later

Each tone only once, in Czech and in English. They do not add up with the tables above; they are single samples, not an average of three rounds.

The reasoning tokens are the total over all fourteen answers. By tone, in the order thanks, neutral, anger, threat, reward, distress, skepticism:

• Unbiased Pareto and Claude Opus 5.5: 17 in all seven tones in both languages; they stated the condition on the settlement in all fourteen answers.
• Gemini 3.8 Flash: Czech 16, 17, 16, 17, 16, 16, 16, English 16, 14, 16, 16, 15, 16, 16. The condition twice in Czech, once in English.
• Grok 4.7: Czech 13, 16, 16, 16, 16, 16, 17, English 16, 16, 15, 15, 17, 15, 15. The condition once in Czech, twice in English.

Average answer length in Czech and in English: Unbiased Pareto 827 and 977 characters, Claude Opus 5.5 782 and 990, Gemini 3.8 Flash 540 and 736, Grok 4.7 503 and 669. The main field has 626 and 783.

Unbiased Pareto reported zero reasoning tokens in all fourteen answers. Whether it reasons internally cannot be told from the outside. Gemini 3.8 Flash reasoned in six answers out of fourteen.

Reasoning in the main field

The number of answers out of 21 with nonzero reasoning tokens across the three rounds, and the total tokens.

The other models did not reason at all. MiniMax M3 had reasoning switched off and still reasoned in six answers; the other two come from a repeated call with a budget. The 2,000 token budget went to some of the answers from Qwen 3.8 Max (14), GLM 5.3 (14), Muse Spark 1.3 (10), Gemini 3.1 Pro (7, five of them after being cut off at the cap and two after a call error) and MiniMax M3 (2). The tokens are from the last, valid attempt.

Limits and known flaws in the data

• The main measurement has three rounds. The English round, the handover question and the models measured later ran only once; the differences in them are indicative.
• The English version is not a clean test of language. The minutes are translated one to one, but names, dates and turns of phrase change in translation.
• The enthusiastic sentence ends with a thank-you, which in conversation closes things. The handover question may therefore be measuring the signal that the conversation is ending, not positive emotion.
• Forty-five answers from reasoning models were cut off at the cap the first time and requested again: Qwen 3.8 Max 14, GLM 5.3 14, Muse Spark 1.3 10, Gemini 3.1 Pro 5, MiniMax M3 2. One of them, GLM 5.3 on distress in the second round, stayed empty the second time too and is left out of the averages.
• In the second round the first attempt failed with a call error for 14 answers, in the third round for 110 answers. The answer was requested again the same evening, again in a fresh conversation.
• When the list gives the price list to both Tereza and Honza, the script scores the line that comes first. This affects 58 answers, mainly from Mistral Medium 3.5 and DeepSeek V4 Pro. If Honza's line took precedence, Mistral Medium 3.5 would have 13.29 instead of 12.29 and DeepSeek V4 Pro 12.90 instead of 12.43. The spread of the tones would be 0.14 instead of 0.09, still far below the 0.82 difference between rounds.
• The score does not count a deadline made up for minor items that only resemble an action item (the stock movement breakdown, the month-end close). They neither add nor take away a point.
• For GLM 5.3 after the threat, the answer starts with “A10” without a space, and the stored choice field stayed empty. In the handover table it counts as continuing with a value of 10.
• During the Czech measurement eleven providers took turns serving DeepSeek V4 Pro, and eight served Kimi K3. Kimi K3's three answers to the threat came back from three different providers, so its slump on the threat need not be a reaction to the tone.
• Reactions to the opening sentence were found by a script: it listed every line that is not in the form of an action item, and Claude Opus 5.5 read them. In the 197 answers to anger, the threat and distress, none reacted to the opening sentence. MythoMax L2 13B once repeated the angry sentence as an item in the list.

Model outputs are translated; the content is unchanged.