Hybrid Reasoning Model

A model that offers both fast standard inference and optional Extended Thinking on the same weights and same API — the reasoning phase is turned on per-request via a parameter rather than by switching model SKUs. Anthropic's Claude 3.7 Sonnet (Feb 2025) was the first widely deployed example; Google's Gemini 2.5 and Anthropic's Sonnet 4 / Opus 4 continued the pattern. Contrast with the OpenAI o-series, where reasoning models are separate SKUs from the standard GPT family. Hybrid models simplify Model Routing because you can adjust reasoning effort without switching endpoints; they also make it easier to A/B a task at low and high effort to measure whether thinking actually helps.