“DeepSeek's R1 model achieved a 35x cost reduction for input-output tokens compared to OpenAI's reasoning models.”