Full definition
A reasoning model (OpenAI o-series, DeepSeek-R1, etc.) produces an internal monologue, a chain of thought of invisible tokens, to break down a problem before replying. This multiplies generated tokens (often by 3 to 15) and therefore energy: from a few Wh for a small model to more than 30 Wh on a long prompt. Useful for complex problems, oversized for everyday tasks.