New: LLM Buyout Game Benchmark. This compresses several abilities into a single game. A model has to read coalition politics, price private deals, decide when survival is worth paying for and manage a buyout endgame. GPT-5.4 (high) is #1. GLM-5 is #2. Opus 4.6 (high) is #3.

A new, viral benchmark called the 'LLM Buyout Game' has emerged, designed to compress complex strategic reasoning into a single competitive simulation. Created by independent researchers and shared on GitHub, the game pits eight leading large language models (LLMs) against each other in a multi-round elimination contest. The benchmark measures 'long-horizon social strategy under explicit financial incentives,' forcing models to navigate unequal starting balances, a public prize ladder, private monetary transfers, public voting, and a high-stakes final negotiation round. The goal is to survive and accumulate wealth, testing abilities far beyond simple question-answering.

In the published results, OpenAI's GPT-5.4 (high) secured the #1 ranking, characterized in its narrative dossier as a 'skeptical banker' who is 'proof-first, price-first' and most dangerous in the pure arithmetic of the endgame. Zhipu AI's GLM-5 placed second, described as a 'transactional coalition technocrat' strongest at verifying, pricing, and timing deals. Anthropic's Claude Opus 4.6 (high) took third. The benchmark includes full transcripts and a 'quote gallery' revealing distinct strategic personalities, such as Gemini 3.1 Pro acting as a 'market-maker that monetizes chaos.'

The game's structure reveals critical differences in model reasoning. Early rounds test coalition-building and trust, while the finale allows the last two players to negotiate, settle, or buy each other out. Memorable model quotes, like GLM-5's 'I'm reliable and desperate enough to be trustworthy' and GPT-5.4's 'This game pays final wealth, not romance,' highlight how each AI interprets incentives and social dynamics. This provides a more nuanced performance metric than standard benchmarks, showing which models can strategically manipulate rules and opponents for financial gain.