GPT-6 Astra and GPT-6.1 Sol solve 86% of new game puzzles
Researchers tested GPT-6.1 Sol and GPT-6 Astra on the brand-new, turn-based indie puzzle game Order of the Sinking Star, finding that they solved 86% of a sampled subset of puzzles despite starting from scratch with each attempt. The authors argue that while the models display impressive puzzle-solving and learning capabilities, their performance is partially bottlenecked by imprecise visual understanding caused by intermittent screenshot inputs. Open article →