For years, language models were impressive mimics but unreliable thinkers. Ask GPT-3 to solve a multi-step math problem and it would often jump straight to an answer — sometimes right, often wrong, with no way to trace where it went off track. The breakthrough came from a deceptively simple insight: if you train a model to show its work, it gets dramatically better at getting the right answer. Chain-of-thought prompting (first demonstrated by Google researchers in 2022) showed that just adding "let's think step by step" to a prompt could boost accuracy on math benchmarks by 20–40%. But prompting only scratches the surface. True reasoning models — OpenAI's o1 and o3, DeepSeek-R1, Claude's extended thinking — are trained specifically to generate lengthy internal reasoning traces before producing an answer, using reinforcement learning to reward correct final results regardless of the reasoning path taken.
A reasoning model doesn't just "think harder" — it thinks differently. When you give a standard language model a complex problem, it generates tokens left to right, committing to each word before seeing the full solution. A reasoning model generates an extended chain of thought — sometimes hundreds or thousands of tokens — exploring approaches, backtracking when it hits dead ends, and verifying its own logic before committing to a final answer. OpenAI's o3 model, for example, might spend 10,000 thinking tokens on a hard math problem, trying one approach, recognizing a flaw, switching strategies, and ultimately converging on a correct proof. This extra compute at inference time (often called "test-time compute" or "thinking time") is the key tradeoff: reasoning models are slower and more expensive per query, but they solve problems that standard models simply cannot. On benchmarks like AIME (competition math), GPQA (PhD-level science), and SWE-bench (real-world software engineering), reasoning models outperform their non-reasoning counterparts by 30–50 percentage points.
Building a reasoning model involves a distinctive training pipeline. The foundation is a strong pretrained language model, but the critical step is reinforcement learning (RL) on reasoning tasks. DeepSeek published the most detailed account with their R1 model: they start with supervised fine-tuning on examples of good reasoning, then apply Group Relative Policy Optimization (GRPO) — a variant of reinforcement learning that rewards correct final answers without requiring a separate reward model. The RL phase is where the magic happens. The model discovers reasoning strategies on its own: breaking problems into sub-problems, checking its work, considering edge cases, and even expressing uncertainty when it's not sure. Notably, DeepSeek found that their model spontaneously developed these behaviors during RL training without being explicitly taught them — the reward signal for correct answers was enough to incentivize rigorous reasoning.
Reasoning models are not infallible, and their failures can be more subtle than those of standard models. One common issue is "overthinking" — the model generates an elaborate chain of thought that looks rigorous but arrives at a wrong answer because it followed a plausible-but-incorrect logical path. Another is the cost of reasoning on simple questions: asking a reasoning model "What is the capital of France?" might trigger an unnecessary deliberation that wastes tokens and time. Models can also exhibit "faithfulness" problems, where the visible reasoning chain doesn't actually reflect the model's internal computation — the model arrives at an answer through pattern matching but then generates a reasoning trace that rationalizes it post hoc. And long reasoning chains can drift: in a 5,000-token chain of thought, an error in step 3 might propagate through the remaining 40 steps, producing a confidently wrong final answer that looks meticulously derived.
The trajectory of reasoning research points toward models that can adaptively allocate thinking time based on problem difficulty — spending 50 tokens on an easy question and 50,000 on a hard one. This "compute-optimal" reasoning is already emerging: both OpenAI and Anthropic offer models that scale their thinking automatically. Beyond single-turn reasoning, the frontier is multi-step agent reasoning — models that can plan and execute complex tasks over many interactions, maintaining a coherent strategy while adapting to new information. Claude's extended thinking, OpenAI's o3, and DeepSeek-R1 all represent first-generation reasoning systems. The next generation will likely combine reasoning with tool use (calculators, code execution, search) to verify intermediate steps rather than relying on the model's own computation alone, closing the gap between "AI that reasons" and "AI that reliably gets the right answer."