The model, called Flash Thinking, uses techniques pioneered by OpenAI’s o1 model, breaking prompts into steps and “thinking over” the answers.