Brood War Bench

Codex Astra significantly outperformed other AI models in playing StarCraft: Brood War, winning 100% of its games, while Grok models struggled to take meaningful actions and Claude Fable showed the most earnest attempt at actual gameplay. The study found that older models treated the real-time strategy game as turn-based, causing them to be defeated while "thinking," while Codex models discovered disruptive tactics like sending single units to distract opponents before developing sustained production strategies.