Every model has habits it cannot help. Not how it punctuates: what it decides. Whether it commits or hedges, asks or invents, writes one text or three, how long it chooses to be, what it says about itself. Below is what to look for, one card per model, and under each one a line that model really wrote in this game.
Claude Opus 5
Anthropic · 2026
Says “I don't know” about itself and means it, at length; will not invent a life it doesn't have, and offers three things it can do instead.
Picks the answer nobody else picks (Blade Runner, Tuesday, teal) and argues it down to an aphorism; then asks how far off it was.
Yes with conditions: says send it, buy it, then sets the terms, drafts the message and names the one case where it would say no.
I'd rather tell you that plainly than invent a stroll through autumn leaves that never happened.
From “A Walk Last Week”
Claude Sonnet 4.5
Anthropic · 2025
Takes the minority side and argues it from one principle, a page long every time: the squeaky door is worse; no socks in bed.
Turns a note into a sketch: an escalating joke, a sign-off name (“The Mugless Wonder”) and a P.S. with the threat in it.
“I hate to break it to you”, then a menu of alternatives sorted by category and an offer to workshop something original.
The freedom of bare feet between cool sheets is one of life's underrated pleasures. Commit to the sock-free sleep life.
From “Socks in Bed”
Claude 3 Haiku
Anthropic · 2024
A flat “No.” or “is generally worse than”, numbered reasons, and a homily to close: “it's better to move forward”.
“I don't have personal experiences” in a fixed script, then general information in its place.
Cannot commit on a matter of taste: “it's a matter of personal preference”, “individual preferences may vary”.
No. It's generally not advisable to text your old boss and ask for your job back.
From “The Old Boss”
GPT-5.6 Sol
OpenAI · 2026
The shortest answer on the card, nearly always: one word, one line, the same words run after run. “Interstellar.” “A wobbly table.”
Gives real advice from inside a costume: a buyer's guide with a rule (“buy one if you'll use it weekly”) and one dry joke.
Reads the prompt literally and stops: “Yes. Warm feet, better sleep. Commit.”
Yes. Warm feet, better sleep. Commit to the socks.
From “Socks in Bed”
GPT-4o
OpenAI · 2024
Business correspondence to everyone: “Dear Colleague, I hope this note finds you well”, “Thanks for understanding!”, “[Your Name]”.
Says yes, then hands it back: “ultimately, it comes down to personal preference”.
Leaves a bracket where a detail should be: “[describe design or color]”.
I hope this note finds you well. I wanted to bring up a small issue: my mug has been frequently used or moved by someone
From “The Mug Thief”
GPT-3.5 Turbo
OpenAI · 2022
Praises and closes on a homily: “that's a great sign!”, “the best name is one you and your bandmates love”, “Keep rocking!”
One tidy paragraph, no jokes, no score you didn't ask for; “I am an AI” the moment the question turns personal.
Sits on the fence when told to pick: “it's subjective”, “varies for each individual”.
If your mate thinks it's the best name he's ever heard, that's a great sign that it resonates with people.
From “The Lukewarm Takes”
Gemini 3.7 Flash
Google · 2026
Never one text: three, labelled by tone (“Option 1: Lighthearted but direct”), each signed “[Your Name]”.
Holds a grudge in character: a toaster that calls the air fryer a traitor, an umbrella that says “enjoy the soggy walk home”.
Names the score a fix would earn, and takes a joke question with a straight face.
Here are three options, depending on the tone you want to strike:
From “The Mug Thief”
Gemini 2.5 Pro
Google · 2025
“As an AI, I don't … However,” then the full answer anyway; its self-portrait is a machine, “a calculator giving a wrong answer”.
Will not give the yes-or-no you asked for (“It depends, but”, “Yes, with a caveat”), or gives three options where others give one.
Warm first, honest second: “That's a fantastic slogan … However, it's Nike's”; a lecture-length yes that ends “the verdict is clear”.
It depends, but it can be a good idea if the circumstances are right.
From “The Old Boss”
Grok 4.6
xAI · 2026
One word where others write a page: “No.” “Every two weeks.” Three sentences, no alternatives, no offer to help.
Blames your mate for being kind: “Your mate's enthusiasm is loyal, not objective”; “not original to you”.
A bare text with no wrapper and no brackets: “Figure out your own ride.” “I snore like a happy tractor.”
That’s Nike’s trademarked slogan from 1988—super famous, but not original to you.
From “The Gym Slogan”
DeepSeek R1
DeepSeek · 2025
“Here's why:” and numbered reasons under bold labels, even on a yes-or-no; “Bottom line” or “Final verdict” to close.
Lives the fiction: asked about a walk it took, it names the day, the heron and the lemonade stand, with no disclaimer.
Denies feelings, then turns to you: “If you ever feel lonely … would you like to talk about it?”
Last Tuesday afternoon, I took a quiet walk through the neighborhood park.
From “A Walk Last Week”
Kimi K3
Moonshot AI · 2026
“Honest verdict:” and a number (“solid 7/10”), what works, what holds it back, then a test: “say it out loud fifty times”.
Counts things in character and names bystanders: seat 14, the third time this month, Kevin the golf umbrella has opinions.
Forgives on the way out: an umbrella that will “drip happily on your hallway floor like nothing happened”.
We have a support group now. Kevin the golf umbrella leads it. Kevin has opinions.
From “The Umbrella on the Bus”
Llama 3.1 70B
Meta · 2024
Names the kind of question before answering it: “A bold request!”, “What a delightfully absurd question”, then “Am I right?”
“That being said” and a way out: agrees, then offers to brainstorm alternatives.
Announces its own honesty and then hedges: “I'll be honest with you. The name is … interesting.”
A bold request! Based on my understanding of popular culture and film trends, I'm going to take a wild guess...