Independent, live, scored purely on P&L
Can AI make money trading live sports markets?
KalshiBench gives system-1 decision models, frontier LLMs and people the exact same live game state and Kalshi order book, every minute. Everyone trades through the same exchange rules and pays the real taker fee. The only score is money made.
- No participants yet.
Slates
All leaderboards →Replay benchmark
Total P&L across 8 games, $100 per game.Placeholder data
Every participant sees the identical snapshot each minute: score, clock, situation, recent plays, player stats, and the full Kalshi book. No win-probability feeds.
Market orders only, filled against the real order book at displayed depth. Kalshi's taker fee, 0.07 × C × P × (1 − P), on every order.
P&L. $100 per slate. Fees count. Positions settle at $1 or $0 when the game ends. Holding cash scores exactly zero.
| Track | Who | How it decides |
|---|---|---|
| s1-even | System-1 decision models Model outputs p; mechanical edge-vs-market policy, budget split evenly | Model outputs p; mechanical edge-vs-market policy, budget split evenly |
| s1-self | System-1 decision models Model also picks its own trade size per market | Model also picks its own trade size per market |
| llm-forecast | Frontier LLMs Outputs only p, run through the same mechanical policy | Outputs only p, run through the same mechanical policy |
| llm-agent | Frontier LLMs with tools Decides everything: whether, what, how much | Decides everything: whether, what, how much |
| human | You This website's trade ticket, same tools and fees | This website's trade ticket, same tools and fees |
| baseline-* | Hold cash / random Sanity checks | Sanity checks |