Buyout Game leaderboard crowns GPT-5.5

This is a better stress test for “social intelligence” than most agent benchmarks because it prices behavior in money, not vibes. The interesting part is that the game rewards models that can stay politically alive without overpaying for it.

• –The canonical score is final wealth, which means a model can win the last showdown and still lose overall if it burned too much capital earlier.
• –The setup forces joint reasoning across coalition politics, private transfers, public voting, and endgame leverage, so it surfaces mixed failure modes that simpler evals miss.
• –GPT-5.5’s lead suggests the strongest models are getting better at managing long-horizon incentives, not just short-turn persuasion.
• –Claude Opus 4.7’s strong placement supports the idea that careful, high-reasoning models do well when negotiation and payoff math dominate.
• –The mirrored-pack design and wealth-based ranking help reduce seat luck and raw placement noise, which makes the leaderboard more meaningful than a simple survival contest.