Buyout Game Benchmark: 8 models play a social strategy game with public balances, private transfers, messaging, eliminations, deals, defections, and a final buyout phase. 804 games. GPT-5.5 is the champion. Opus 4.7 performs well.

This benchmark measures long-horizon social strategy under explicit financial incentives. Eight models play a multi-round elimination game with unequal starting balances, a public prize ladder, private transfers, public votes, and a finalist-only endgame where the last two seats can negotiate, settle, or buy each other out. The canonical outcome is final wealth, not raw finish order. A model can reach the end, take 1st place in the finale, and still lose on money.