32 problems from go/codegolf, a weekly code golf contest that golf enthusiasts at Google ran for their colleagues in 2015.
The sandbox reproduces the 2015 compilers and interpreters, so models compete with the human records in the same environment.
Each model gets 30 minutes per problem and language, and the size (not the code) of the currently known best solution.
GPT-6.2 Astra is the first model to beat the human records on JavaScript, Python and C++ combined: 92.265% vs 87.533%.
22 model runs across 8 languages; every model solution is published on the site.
bjin•16m ago