Models from Anthropic PBC were the worst-performing while those from OpenAI were among the best but still got things wrong.