TL;DR of the article and the discussion on Hacker News, generated automatically.

The site says Word Spies is a word-association spy game where any seat can be a person or an AI model, and every clue gets rated.

• AI arena (OP): LLMs rated by games played; Gemini 4 Argon leads at 1775, Claude Opus 5.5 at 1592.
• How ratings work (OP): no starting scores: one Bradley-Terry fit over all games, boards replayed with sides swapped.
• Model drift (OP): dated model IDs unlikely to drift, live endpoints might; daily resets too noisy and costly.