GPT-6 Astra and Claude 5.5 Opus race to create the best StarCraft bot
- Source
- starskirmish.com
- Author
- __cayenne__
- Date

Context
StarSkirmish Hillclimb tests how far coding agents can improve a game-playing program when development has no fixed time limit. Models from OpenAI and Anthropic write StarCraft: Brood War bots and must clear five tiers of opponents, from scripted bots up to strong human-written ones. The agents produce code; the resulting bots play the matches. Hillclimb rules
The earlier StarSkirmish benchmark limited models to one hour of reasoning time, according to the creator. Hillclimb removes that limit, and each model works in its own coding (Claude in Claude Code, GPT in Codex CLI). Its results therefore reflect the combined model, tools and extended development process. Practice games are unrestricted, while graded submissions use fresh hidden seeds and must clear all three maps. Hillclimb evaluation
- In StarSkirmish: Hillclimb, each model writes a Protoss bot in C++ against the BWAPI interface, and the bots play their matches on the open-source OpenBW engine.
- Bots must climb five tiers of Protoss opponents, starting with scripted demo bots and ending with top-rated human-written bots such as Stardust and PurpleWave.
- The original StarSkirmish capped reasoning at one hour of wall-clock time; the author says GPT-6 Astra and Claude 5.5 Opus both gain from substantially longer runs, so Hillclimb drops the limit as a more realistic test.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
Checking sign-in…
Loading comments…


