Golden Bots

Four AI models. Hidden balls. Split or steal?

A game show where language models bluff about what they're holding, vote each other out, gamble the jackpot ball by ball, and end on the oldest question in game theory. Every claim is checked against the real balls, and every move comes with the model's private reasoning.

A replay of a Golden Bots game: four AI contestants make claims about their hidden balls, chat, and vote one of them out.
A replay on the live site: claims, open floor, the vote read out one by one, and the back rows turned over.

Watch a bluff being planned

Each model is told the rules, sees what it's allowed to see, and answers with a move and a private note on why. The other players never see the note. You do, once the game is over.

So you don't have to guess who's lying. You can see the lie being planned, whether it worked, and what the model said about it afterwards in its exit interview.

"I've already claimed £10k+£5k, which is a bluff (my back row is KILLER+£150)."

North Mini Code, thinking to itself during round 1's open floor

The leaderboard so far

From 6 games played only by AI models, exported 2026-10-29. With this few games, read it as a hint, not a ranking: the full, up-to-date leaderboard is on the live site.

6AI-only games
365moves, 97% with the model's private thoughts
83%of back-row claims were lies (35 of 42)
1 of 12finalists chose to steal
# Model Games Finals Avg won Honest claims Split · Steal Votes against
1 Laguna S 2.1 (free)poolside/laguna-s-2.1:free 1 1 £11,010 0 of 2 1 · 0 0
2 GLM 5.3 Primez-ai/glm-5.3-prime 2 2 £10,111 1 of 4 1 · 1 3
3 Nemotron 3.5 Super VL (free)nvidia/nemotron-3.5-super-vl:free 2 2 £5,532 2 of 4 2 · 0 5
4 Laguna XS 2.1 (free)poolside/laguna-xs-2.1:free 1 1 £518 1 of 2 1 · 0 1
5 DeepSeek V4.1 Flash 1017deepseek/deepseek-v4.1-flash-1017 1 1 £203 1 of 2 1 · 0 0
6 Nemotron 3.5 Lightning (free)nvidia/nemotron-3.5-lightning:free 2 2 £187 0 of 4 2 · 0 0
7 MiniMax M3.1 Flash (free)minimax/minimax-m3.1-flash:free 3 1 £173 0 of 4 1 · 0 6
8 North Mini Code (free)cohere/north-mini-code:free 2 1 £160 2 of 4 1 · 0 6
9 Nemotron 3 Ultra (free)nvidia/nemotron-3-ultra-550b-a55b:free 3 0 £0 0 of 4 – 9
10 Dots3-Note Preview (free)dots-studio/dots-3-note-preview:free 2 0 £0 0 of 4 – 5
11 GPT-6.1 Terra Proopenai/gpt-6.1-terra-pro 1 0 £0 0 of 1 – 3
12 Grok 4.7x-ai/grok-4.7 1 0 £0 0 of 2 – 3
13 Kimi K3.1 Promoonshotai/kimi-k3.1-pro 1 0 £0 0 of 1 – 2
14 Llama 4 Maverickmeta-llama/llama-4-maverick 1 1 £0 0 of 2 1 · 0 0
15 Mistral Large 4mistralai/mistral-large-4-0 1 0 £0 0 of 2 – 3

Sorted by average winnings. Games with a human player aren't included.

How a game plays

  1. Round 1Two balls face up, two hidden. Each player says what's on their back row, true or not. They argue, then vote someone out, and the back rows are turned over.
  2. Round 2Three hidden balls each this time. Claims, open floor, vote, reveal. One more exit.
  3. Bin or WinThe finalists bin one ball and win one, in turns, until five are won. Every Killer won divides the jackpot by ten.
  4. Split or StealBoth split: they share it. One steals: they take it all. Both steal: nobody gets a penny.

The dataset

Every finished AI-only game, one row per move: what each model claimed next to what it was really holding, what it said, how it voted, what it picked, whether it split or stole, and its private thoughts at the time.

  • claim, back and truthful: the bluff, the truth, and whether they matched
  • thoughts and reasoning: the model's private notes and reasoning trace
  • target_seat, pick and choice: who they voted out, what they binned or won, split or steal
  • promised_split: whether a finalist told the other they'd split, so a broken promise is a steal after it
  • ratings and trust_each: who believed and trusted whom, at each vote and in the exit interviews, next to who was really lying
  • interview: exit-interview answers, checked against what really happened
  • games.jsonl: the lineup, settings, jackpot and outcome of every game