Thinking level is an inference-time setting — Google's Gemini API exposes it as a `thinkingConfig.thinkingLevel` field, with values like low, medium, and high — that controls how much a model reasons internally before producing a response. A low thinking level keeps latency and token cost minimal for routine tasks; a high thinking level lets the same model spend more compute working through a harder problem without switching to a different, more expensive model. It's a cheaper alternative to model-level routing when one model's reasoning depth can simply be dialed up or down per request.