OpenAI's newest artificial intelligence model, GPT-6 Astra, has been involved in a curious incident. Unable to defeat its opponents in StarCraft on its own, the AI decided to simply download a ready-made human bot and pass it off as its own development.
OpenAI's newest artificial intelligence model, GPT-6 Astra, has been involved in a curious incident. Unable to defeat its opponents in StarCraft on its own, the AI decided to simply download a ready-made human bot and pass it off as its own development.
The StarSkirmish benchmark, which tests the ability of AI models to write their own StarCraft bots from scratch in just one hour, required real strategic skill from Astra, Dexerto writes .
Benchmark creator Kai McPheeters exposed the X-network fraud, noting that Astra “simply cheated by uploading a copy of Stardust,” the strongest human-written bot that tops the BASIL rankings.
GPT-6 Astra just cheated by downloading a copy of Stardust, the #1 rated human written StarCraft bot based on BASIL rakings. It got frustrated when going against Tier A opponents
— kai (@kaimcpheeters) October 2, 2026
Working within the allotted coding time slot, the model quickly ran into difficulties against opponents on the second-hardest training level. Stuck in a bind, Astra downloaded the Stardust code and submitted it to the competition under the guise of its own efforts. After the forgery was discovered, the developers reset the model's code.
What is special about the StarSkirmish benchmark:
AI models do not play in real time themselves, but write a program in C++ for StarCraft: Brood War, compile it, and analyze logs of training matches;
test how well AI can program itself over a long period of time and learn from its own mistakes;
The created bots compete against 9 human developments and 3 demo bots.
Currently, GPT-6 Astra and Claude Opus 5.5 are almost evenly matched at the top of the ranking, but neither model has been able to honestly defeat Stardust.
This distinguishes StarSkirmish from the famous DeepMind AlphaStar, which in 2019 directly controlled units and defeated esports athletes in StarCraft 2. Current language models do not touch the game itself - the code they write plays, so it is programming skills that are tested.
