Brood War Bench

	Brood War Bench(bw.swerdlow.dev)
	338 points by benswerd 24 days ago \| 153 comments
	tl;dr: A new benchmark pits LLM agents against each other in StarCraft: Brood War, with Codex Astra (xhigh) sweeping 18-0, followed by Claude Fable, while Grok 4.6 struggled to even produce combat units—one game logged 11K reasoning tokens but only 6 command batches in 43 minutes. None of the models played beyond beginner level: a human doing a basic photon rush would beat them all. Older models tended to treat the RTS like a turn-based game and got destroyed while "thinking."
	HN Discussion: •Nostalgic reminiscence about playing StarCraft in earlier eras, unrelated to the benchmark itself ↑The benchmark is valuable for quantifying model speed and real-time decision-making tradeoffs ~Surprised at poor LLM performance given past AI successes in games like DOTA and older Brood War AI tournaments ↑Strategy games are a generally valuable domain for evaluating LLM capabilities •Appreciation for the site's technical implementation (replay viewer) rather than the benchmark findings