Show HN: Gemini 4 Argon plays StarCraft
Points and comments are a snapshot, not live.
StarSkirmish Bench v0.2 tests LLMs by having them write StarCraft bots in one hour.
The StarSkirmish Bench evaluates how strong a StarCraft: Brood War Protoss bot each LLM can write in C++ given one hour. GPT-6 Astra scored 57, followed by GPT-6.1 Sol at 46, then Gemini 4 Argon and Claude Opus 5.5 tied at 39. Against top human-written bots (Stardust, PurpleWave, BananaBrain), even Astra wins only 7% of games. Gemini 4 Argon preferred one-shot bot writing and practiced against high-tier opponents early, but its bots won only 7% of those games. The strongest predictor of score is how much a model practices against strong opponents.
What commenters are saying
Commenters were impressed by Astra's 57% win rate but noted the drop to single digits against top human-written bots, which one commenter explained is because those bots are generally harder opponents. Some discussed the Pluto bot, a neural network bot that beats top human players but has exploitable weaknesses like no memory of units out of sight. Others questioned how LLMs avoid relying on pre-training data and noted the advantage of the one-hour time limit. One commenter detailed how Astra adapted strategies over time, switching to a proxy 2-gate Zealot rush against stronger bots.