Reasoning AI Models: How They Work and Why They Matter

In the fall of 2024, OpenAI introduced the o1 model, which became a sensation: on the AIME math exam it scored 74% in single-attempt mode, and up to 93% with consensus selection, whereas GPT-4o managed only 12% of the problems. On the GPQA Diamond benchmark, o1 achieved 78%, surpassing the average expert. These results demonstrate the capabilities of Large Reasoning Models (LRMs), which emulate the slow "System 2" thinking described by Kahneman. Other models classified as LRMs include Anthropic's (Claude Fable 5, Mythos 5, Opus 4.8), Google's Gemini 3.1 Pro, DeepSeek R1, and GLM-5.2. Unlike classical autoregressive large language models (LLMs), which operate like the fast "System 1," reasoning models use chains of thought, additional computation at inference time (test-time compute), and reinforcement learning. They generate reasoning tokens that are invisible to the user, which increases request processing time to a minute or more. Research has shown that the "think longer" strategy is more effective than simply scaling up model size: a model can perform as well as an LLM 14 times larger if given enough time to think. China's DeepSeek-R1 integrated reasoning directly into the model itself, doing away with an external "critic."