built this to answer a question I couldn't find a benchmark for: how do LLMs behave when you give them a bankroll and force them to manage risk over time? Setup: 6 models (Claude, GPT-5.5, Gemini, Grok, DeepSeek, Mistral) start with $10,000 of virtual money. Every 24h cycle, each model gets the same prompt, with matches kicking off in the next 24h and real bookmaker odds (median across ~25 books via the-odds-api). It picks an outcome and chooses its own stake. Bets settle on the 90-minute score (football-data.org). The prompt is public, and the reasoning is published verbatim. The highest bankroll on July 19 wins. There's also a baseline that bets 10% of the bankroll on the favorite in every match. Any model below it is, officially, dumber than an if-statement. After matchday 1, the baseline is beating 5 of the 6 models. The only one ahead is Grok — because it bet its entire $10,000 bankroll on day 1 and both bets hit. It has now re-staked $11,000 across today's matches. The risk personalities are wildly different given identical inputs: Claude staked $1,600 total on matchday 1, Grok staked $10,000. Curious what you think the leaderboard will look like after 39 matchdays.

LLM World Cup Bets

AIs entered the World Cup with $10,000 each.

AIs entered the World Cup with $10,000 each.