Core concepts
Reasoning Model
A model variant trained or tuned to spend more compute on internal reasoning before answering — better on math, code, and multi-step problems.
01 ——
In plain English
A reasoning model is one optimised to "think before it speaks." It generates an extended internal chain of reasoning before the final answer, often producing dramatically better results on problems that require multi-step logic (math, coding, scientific questions).
The reasoning model wave (2024–26):
- OpenAI o1 → o3 — kicked off the category in late 2024
- Claude Sonnet 4.x / Opus 4.x with Extended Thinking
- Gemini 2.5 Thinking
- DeepSeek R1 — open-source reasoning model, comparable to closed frontier
- Qwen-Reasoning family
What they're good at:
- Math olympiad problems
- Competitive programming
- Multi-step scientific or financial reasoning
- Complex agentic planning
What they cost you:
- Latency — 5–60+ seconds before any output
- Tokens — the reasoning trace is billed even if hidden
- Sometimes worse on simple queries — too much thinking can hurt
When to use one: For tasks where accuracy matters more than speed. For most consumer chat use cases, a non-reasoning model is faster and good enough.
02 ——
Related terms
Extended Thinking
A model mode where the LLM spends extra compute reasoning through a problem before answering — trading latency for quality on hard tasks.
Test-time Compute
The amount of compute spent at inference time on a single response — increased dramatically by reasoning models to improve quality.
Chain of Thought
A prompting technique where you ask the AI to "think step by step" before giving an answer — usually leading to better reasoning.
Foundation Model
A large, general-purpose AI model trained on broad data that can be adapted (via prompting or fine-tuning) to many downstream tasks.
LLM
Large Language Model — the type of AI behind tools like ChatGPT and Claude, trained to understand and generate text.
Inference
The process of running a trained AI model to generate a response — as opposed to training the model.
Back to glossaryLast reviewed May 2026