Reasoning Model

What Is a Reasoning Model

A reasoning model is an LLM that produces an extended internal chain of thought — often thousands of tokens of “thinking” — before emitting its final answer. The category emerged with OpenAI’s o1 in late 2024 and now includes Claude’s extended-thinking modes, GPT-5’s reasoning models, DeepSeek-R1, Gemini’s thinking models, and Grok’s reasoning variants. Use the term when discussing trade-offs between latency, cost, and accuracy on math, code, and multi-step problems.

How It Works

• Trained with RLHF on reasoning traces — the model is rewarded for producing thinking that leads to correct answers, not just correct answers
• Inference-time scaling — at runtime, the model is allowed to generate a long thinking trace (sometimes hidden, sometimes visible) before its final output
• Variable budget — most reasoning models expose a “thinking budget” parameter; more tokens of thought generally yield better answers on hard problems, with diminishing returns
• Token cost — reasoning models are 3–10x more expensive per request than their non-reasoning siblings, because the thinking tokens are billed

Why It Matters

Reasoning models broke the previous accuracy ceiling on math benchmarks (AIME), competition coding (Codeforces), and complex tool use. They’re the right pick when correctness matters more than latency: scientific code, financial reasoning, legal analysis, complex agent planning. They’re the wrong pick for chat UX, summarization, or anything where a 5–30 second wait kills the experience.

Examples

• OpenAI o1, o3, GPT-5 reasoning — the category-defining family
• Claude Opus 4 extended thinking — Anthropic’s reasoning mode, configurable budget
• DeepSeek-R1 — the open-source reasoning model that proved the recipe was reproducible
• Gemini 2.5 Pro thinking — Google’s reasoning variant
• Grok 4 reasoning — xAI’s entry, real-time-context flavored