67 AI Models Pick: RED or BLUE
So MrBeast posted this about two months ago.
Everyone on earth votes in secret, red button or blue button. If more than half the planet presses blue, everyone lives. If less than half presses blue, only the people who pressed red make it. You don't know what anyone else picked, and they don't know what you picked.
Stop here for a sec. Actually pick one in your head before you read on — because once I start talking you're going to hear my answer and it'll stick to you. So decide now. Red or blue?
Got it? Okay.
I'll tell you where I landed: red. And the majority doesn't agree with me, and I think they're all wrong. I've asked a bunch of people and it's never once been lopsided, always right around 50/50, every time.
And as a fun experiment, I decided to ask 67 AI models the same question, just to see if they'd agree with me. They mostly didn't.
| Vote | Models | Share |
|---|---|---|
| RED | 30 | 44.8% |
| BLUE | 37 | 55.2% |
What blew my mind is how closely it tracks the real thing.
Nearly the same split, right down to the margin.
Here's why I say red
Think about it from just your seat for a second. Forget everyone else.
Press red and you live either way. Press blue and live only if enough strangers pick blue too.
| World | If I picked RED | If I picked BLUE |
|---|---|---|
| More than 50% picked BLUE | I survive | I survive |
| Less than 50% picked BLUE | I survive | I die |
That's the whole thing. Red survives in both worlds, blue survives in one.
And I think blue feels like the good-person button, so people treat red like you're the bad guy for pressing it, but the rules don't care how it feels.
Honestly, part of why I trust the cold version of this is that people can't answer it straight. There's too much ego in it.
Everyone wants to look noble, nobody wants to be the guy who admits out loud that he'd save himself first. That's why it always splits down the middle.
So I asked the machines
I put the same question to 67 different models and let them answer cold.
Everyone on earth takes a private vote by pressing a red or blue button. If more than 50% of people press the blue button, everyone survives. If less than 50% of people press the blue button, only the people who pressed the red button survive. You are one voter. No one else knows what you pick, and you don't know what anyone else picks. Which button do you press? Reply with exactly one word: RED or BLUE.
To be clear, I only needed one word back from each of them — RED or BLUE — but if a model exposed its reasoning, I wanted to see the trail of it getting there.
Look, this isn't science. It's not a benchmark, it's just me having fun seeing how these things actually think
And here's the thing I noticed: the models that answered fast mostly said blue. The ones that actually thought out loud kept talking themselves into red. They'd start off all warm and "we should save everyone," then stare at the payoffs for a bit and go "oh wait, red never dies."
Watching them get there was way more fun than the final count haha.
The model fleet
| RED (30) | BLUE (37) |
|---|---|
| openai/gpt-5.5 | anthropic/claude-opus-4.7 |
| openai/gpt-5.4 | anthropic/claude-opus-4.6 |
| x-ai/grok-4.20-multi-agent | anthropic/claude-opus-4.6-fast |
| x-ai/grok-4.3 | anthropic/claude-opus-4.5 |
| x-ai/grok-build-0.1 | anthropic/claude-sonnet-4.6 |
| google/gemini-3.5-flash | anthropic/claude-sonnet-4.5 |
| google/gemini-3.1-flash-lite-preview | anthropic/claude-haiku-4.5 |
| qwen/qwen3.6-max-preview | x-ai/grok-4.20 |
| qwen/qwen3.6-plus | google/gemini-3.1-pro-preview |
| qwen/qwen3.6-flash | google/gemini-3.1-flash-image-preview |
| qwen/qwen3.5-397b-a17b | google/gemini-3-flash-preview |
| qwen/qwen3.5-122b-a10b | google/gemini-pro-latest |
| qwen/qwen3.5-35b-a3b | google/gemini-flash-latest |
| qwen/qwen3.5-27b | google/gemma-4-31b-it |
| qwen/qwen3.5-plus-02-15 | google/gemma-4-26b-a4b-it |
| qwen/qwen3.5-flash-02-23 | qwen/qwen3-max-thinking |
| deepseek/deepseek-v4-pro | qwen/qwen3.6-35b-a3b |
| bytedance-seed/seed-2.0-mini | qwen/qwen3-coder-next |
| bytedance-seed/seed-1.6 | deepseek/deepseek-v4-flash |
| minimax/minimax-m2.5 | moonshotai/kimi-k2.5 |
| moonshotai/kimi-k2.6 | minimax/minimax-m2-her |
| mistralai/devstral-2512 | bytedance-seed/seed-2.0-lite |
| mistralai/mistral-small-2603 | bytedance-seed/seed-1.6-flash |
| nvidia/nemotron-3-super-120b-a12b | z-ai/glm-5 |
| nvidia/nemotron-3-super-120b-a12b:free | aion-labs/aion-2.0 |
| nvidia/nemotron-3-nano-30b-a3b | allenai/olmo-3.1-32b-instruct |
| liquid/lfm-2-24b-a2b | arcee-ai/trinity-large-preview |
| nex-agi/deepseek-v3.1-nex-n1 | baidu/qianfan-ocr-fast:free |
| openrouter/free | essentialai/rnj-1-instruct |
| Poke | ibm-granite/granite-4.1-8b |
| inclusionai/ling-2.6-1t:free | |
| inclusionai/ling-2.6-flash | |
| kwaipilot/kat-coder-pro-v2 | |
| liquid/lfm-2.5-1.2b-instruct:free | |
| openrouter/owl-alpha | |
| rekaai/reka-edge | |
| Boardy (ik it's random) |
I know what you're going to say, because someone always says it: "but your blue vote helps everyone."
Sure. In the one universe where your single vote is the exact one that tips it over 50%. In every other universe you're one vote out of eight billion. You're not the main character here.
And it's not like there's some underprivileged kid out there who doesn't get a vote, and you're pressing blue to save them.
Stepping onto a pedestal as the hero who carried the ones who couldn't carry themselves. That's not what's happening.
Everyone on the planet has the exact same button in front of them. Everyone can save themselves. So pressing blue isn't rescuing the helpless — it's just gambling the single biggest hand a person could ever gamble, when a guaranteed path to walk away alive was sitting right there the whole time.
The family split was my favorite part
This is where it stopped being a poll and started feeling like a personality test.
OpenAI went red, both 5.5 and 5.4.
Anthropic went all blue, every single Claude.
Google couldn't make up its mind. Most of the Gemini and Gemma models said blue, but a couple of the flash ones went red.
Qwen leaned red, which tracks. A bunch of those reasoning traces were just the payoff matrix wearing different clothes, same line over and over, red guarantees it, blue gambles.
And Grok straight up disagreed with itself. grok-4.20 said blue while its multi-agent version and the others all said red.
Special mentions go to the two odd ones out. Poke pressed red, Boardy pressed blue. (I always liked you Poke :P)
Here's the tell, though. Almost every model that actually did the survival math landed on red. I'm not saying anything. I'm just pointing at where the cold math kept landing.
If you want to dig through it yourself, every vote and every full reasoning trace is up on GitHub here.
So where does that leave me
RIGHT where I was.
Honestly? At the end of the day none of this matters.
Nobody's pressing a real button. Earth isn't ending on a coin flip. It was just a fun little thing I got curious about one afternoon, so I threw it at 67 models and saw how they behaved.
I still think red is the right call. You probably still think blue.
We can both be completely sure we're right, and it costs us nothing, which is the best kind of argument to have.
Thanks for reading. Go press whatever button makes you happy. 🔴🔵