{
  "id": 14978741,
  "title": "Show HN: Gemini 4 Argon plays StarCraft",
  "url": "https://urgent.news/2026/10/16/show-hn-gemini-4-argon-plays-starcraft",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-16T14:18:44.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://starskirmish.com/bench/v0-2/"
  },
  "original_language": "en",
  "account": "The StarSkirmish Bench updated its StarCraft bot evaluation methodology in v0.2, incorporating flagship models from OpenAI, Anthropic, Google, and Meta. Each model was given one hour of wall clock time to write its StarCraft bot, with the model's score reflecting its win rate against a roster of competitive human-written bots and demo bots. GPT-6 Astra emerged as the top-scoring LLM in the v0.2 benchmark, outperforming GPT-6.1 Sol in second place. Gemini 4 Argon, Claude Opus 5.5, and Claude Fable 5.1 were functionally tied for third place, all scoring between 37 and 39. While GPT-6.1 Sol offered unique value as the only LLM in the most desirable scoring range, the strongest human-written bots still outperformed the models. Against top opponents like Stardust, PurpleWave, and BananaBrain, GPT-6 Astra won only 7% of its games. Google and Meta's new flagships, Gemini 4 Argon and Muse Wonder Full, replaced v0.1's Gemini 3.8 Flash and Muse Spark 1.3 in the rankings. The strongest predictor of a model's score was how much time it spent practicing against the seven strongest bots on the roster. Some models, like GPT-6 Astra, GPT-6.1 Sol, and Gemini 4 Argon, started practicing these bots early into the hour, while their own bots often lost to them. The results varied between runs of the same model, with Gemini 4 Argon exhibiting the most significant variation. Each LLM wrote a Protoss bot in C++ using BWAPI 4.4.0, played on OpenBW, with three ladder maps: Heartbreak Ridge, Benzene, and Destination. The LLM began from a skeleton bot that only had mining capabilities, with access to the source code of three demo bots and two standard bot libraries. The final bot was compiled and played against 13 other bots, including 10 competitive human-written bots and three demo bots, six times on each map, with fresh seeds. Each model's five final bots played against the 13 roster bots, the 10 competitive human-written bots, and the three demo bots, six times on each map, resulting in win rates used to determine the model's StarSkirmish Bench score.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}