CrusaderBench Makes AI Agents Play Stronghold Crusader; GPT-6.2 Sol Tops Board

Hesamation · x · 2026-11-20

Hesamation shared CrusaderBench, an independent benchmark where AI agents play Stronghold Crusader: each model views the game via screenshots, acts with mouse and keyboard, and runs a medieval village for 33-45 game minutes on each of three maps. Score anchors: 0 = doing nothing, 100 = a human player's game (80% net worth, 10% population, 10% popularity).

Key results:

Post-game interviews are entertaining: Claude Fable 5.1 admitted, "I never checked the actual conversion ratios; I reasoned from unit prices and got it backwards."

Related event: CrusaderBench Tests 15 Models on Stronghold Crusader, GPT-6.2 Sol Leads(2 posts)→

Original post →

More from Models

Models channel →