Jargon:Reasoning model
1 min read
Say this at the party
“An AI that thinks a problem through step by step before it answers, which makes it slower but much better at hard questions.”
In plain English
Ordinary chat models answer straight away, a word at a time. That's fine for easy questions and shaky on anything tricky.
A reasoning model takes a moment first. It writes itself a scratchpad of working out, checks its steps and only then replies. In most apps you'll see it as a "thinking" option.
Use it for maths, planning, code and big decisions. For a quick email, it's overkill.
Where it came from
A 2022 Google paper showed that asking models to show their working, called chain-of-thought, made them better at reasoning. OpenAI's o1, in September 2024, was the first big model trained to do it by itself, and DeepSeek's R1 in January 2025 showed the recipe openly. Most top models now have a thinking mode.
See it in action
Ask "Should I take the job with the higher salary or the shorter commute?" A quick model gives you a tidy list. A reasoning model works through your numbers, spots that the commute costs you £3,000 a year, and asks what you value more.
What people get wrong
- The working out you can see isn't always a faithful record of how the AI actually reached its answer.
- Thinking longer doesn't make it right. Reasoning models still make confident mistakes, just fewer of them.
Go deeper
- PaperWei et al., Google, 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsThe paper that showed "show your working" helps AI too.
- ArticleAnthropic, 2025Claude's extended thinkingWhat thinking mode does, and its limits, in plain terms.
- PaperDeepSeek-AI, 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningHow you train a model to think before it speaks.
Where it comes up
Related jargon
Drafted with AI, checked by the AI Council (Claude, ChatGPT and Gemini) and signed off by Phin.