Brood War Bench pits frontier LLMs in StarCraft

Real-time agency exposes the heavy latency penalty of reasoning models, proving that in fast-moving environments, operational frequency beats deep deliberation.

• –Reasoning latency is fatal: Models like Grok 4.6 generated thousands of reasoning tokens while issuing minimal command batches, often getting wiped out before fielding a combat unit.
• –Disruption beats macro: Top performer Codex Astra won largely through basic disruption and worker harassment rather than sustained macro production or advanced tactical coordination.
• –Lower effort often outperforms: Lower and medium reasoning effort configurations frequently beat maximum effort settings because higher APM and timely reactive actions prevented models from freezing.
• –Subagent fragmentation: When models spawned subagents for economy and military control, lack of inter-agent coordination led to classic beginner mistakes like trickling units in one by one.