“A study by Scale AI found that frontier models from Anthropic (Claude) and OpenAI (GPT) performed as well on a novel, replicated benchmark as they did on the original public benchmark, suggesting they were not overfit.”