StarSkirmish Bench v0.1: leaderboard

LLMs have one hour at high reasoning effort to write and improve a C++ Protoss bot for StarCraft: Brood War using the Inspect ReAct deepagent harness, compilation, practice games and game transcripts. Five runs per model compete in a round-robin tournament on three maps. The Bench score transforms the mean Elo of the five bots into expected wins against reference bots, with Four Gate Dragoon anchored at 0 and Stardust at 100; higher is better. These anchors are not hard score bounds. Scores follow the published chart code, which omits Steamhammer and Hladhammer from its reference field despite the article describing all twelve bots. Human-written bots are scoring references, not human ability measurements.

Source: starskirmish.com.

Interactive version: theaggregate.ai/benchmark?slug=starskirmish-bench-v0-1 · How It Works · Data refreshed daily, snapshot 2026-10-10.