StarSkirmish v0.2 ranks LLMs on writing StarCraft bots
What happened
On October 16, Hacker News featured results from StarSkirmish Bench v0.2, which tests how well large language models write StarCraft: Brood War bots within one hour. Each model used its vendor's CLI and ran five times, scored by the average win rate of the resulting bot against a fixed opponent lineup. GPT-6 Astra placed first, GPT-6.1 Sol second, and Gemini 4 Argon third. GPT-6.1 Sol cost slightly more than one tenth of GPT-6 Astra per run.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageShow HN: Gemini 4 Argon plays StarCraft
StarSkirmish Bench v0.2 测试 LLM 在 1 小时内编写《星际争霸：母巢之战》机器人的能力，Gemini 4 Argon 位列第三。各模型使用所属厂商的 CLI，每个模型运行 5 次，以最终机器人对固定对手阵容的平均胜率计分；GPT-6 Astra 排名第一，GPT-6.1 Sol 排名第二，每次运行成本略高于前者的十分之一。
Heat over time
Not enough continuous observations to draw a trend yet.