Codex on GPT-6 Astra won all matches in the StarCraft benchmark

Freestyle co-founder Ben Sverdlov created the Brood War Bench, where 19 model configurations in Codex, Claude Code, and Grok wrappers played 171 matches in StarCraft: Brood War, with Codex on the Astra model at xhigh reasoning level leading with 18 wins and 0 losses.

Compiled by Sergey KostenchukPublished 2026-09-23Updated 2026-09-23

2026-09-23 Research Anthropic

🤖 Codex on GPT-6 Astra won all matches in the StarCraft benchmark

Freestyle co-founder Ben Sverdlov created the Brood War Bench: 19 model configurations in Codex, Claude Code, and Grok wrappers each played 171 matches in StarCraft: Brood War. The leader is Codex on Astra (xhigh): 18 wins and 0 losses at a cost of about $10.54 per game.

🌍 A rare real-time agentic benchmark exposes the weaknesses of specific wrappers: Codex Astra won with early worker attacks while opponents were 'thinking', and its sub-agents were barely coordinated. Claude Fable was the only one to systematically build economy and technology. Grok 4.6 took only 3 matches.

👤 No model reached a level above beginner: they played like in a turn-based game and died while thinking. Match recordings are at bw.swerdlow.dev/report, and you can run your own agent at bw.swerdlow.dev. The 18-game lead may be noise.

Source 1: https://bw.swerdlow.dev/report Source 2: https://bw.swerdlow.dev/

Sources