Gemini 4 Argon plays StarCraft
The gist
Google's flagship AI model Gemini 4 Argon placed 3rd in StarSkirmish Bench v0.2, scoring 37 to 39.
The benchmark tests whether AI can write working game code, not just play games.
How it unfolded
- date unknownStarSkirmish Bench v0.2 is published, evaluating LLMs that write StarCraft: Brood War bots. Gemini 4 Argon places 3rd, scoring between 37 and 39, functionally tied with Claude Opus 5.5 and Claude Fable 5.1.
- Oct 12, 2026Heiga Zen posts referencing Gemini 4 Argon in the context of gaming and AI breakthroughs mastering Atari, Go, StarCraft and SIMA agents.
- Oct 12, 2026Danielle Fong posts that Gemini 4 Argon replicated internal systems in seconds and references 30 hours of playing StarCraft II.