What are AI reasoning models?
An AI reasoning model is a language model that works through a problem step by step before it answers: breaking it down, trying approaches and checking its work. That makes it slower and pricier, but more reliable on hard, multi-step problems.
Updated September 24, 2026
A reasoning model is an AI model that thinks before it speaks. Instead of answering straight away, it works through the problem internally: restating the question, trying different approaches, checking its own steps and dropping dead ends. Only then does it write the answer you see.
How they work
- A hidden scratchpad. The model writes out step-by-step working, often called a chain of thought. Providers such as OpenAI mostly keep this reasoning hidden, but it still uses up tokens in the model's context window.
- Trained to think. Labs train these models with reinforcement learning, rewarding them for reaching correct answers. DeepSeek reported that its R1 model developed habits like double-checking its work on its own during that training.
- More compute at answer time. NVIDIA calls this test-time scaling: spending extra computing power during inference to explore several paths. A hard question can take more than 100 times the compute of a normal answer.
Most major AI labs now offer this: OpenAI's reasoning models, Claude's thinking modes, Gemini's thinking models and DeepSeek-R1.
What they're good at
Anthropic, Google and OpenAI point to the same strengths:
- math and logic problems
- writing and debugging complex code
- multi-step planning and data analysis
- long tasks for AI agents
The tradeoffs
- Slower: thinking takes time, from seconds to minutes on hard problems.
- Pricier: the hidden thinking is billed like visible output.
- Overkill for simple jobs: Google advises little or no thinking for quick lookups or classification.
A short timeline
OpenAI's o1 system card was published in December 2024, and DeepSeek's R1 paper in January 2025 showed an openly released model could reason at a high level too. Since then, adjustable thinking has become a standard feature of leading models.
Quick questions
What is a reasoning model?
An AI model that plans and weighs options in hidden, step-by-step thinking before it answers, which helps on complex, multi-step problems.
How is a reasoning model different from a regular chatbot?
A regular model answers in one pass with no scratch work. A reasoning model works the problem through and checks its steps first, which helps most with math, tricky code and long tasks.
Should I use a reasoning model for everything?
No. For simple questions, lookups or sorting, thinking adds time and cost without much benefit, so a faster setting is usually better.
Why do reasoning models cost more?
Their hidden thinking uses extra tokens, which are billed like the visible answer, and hard questions can take many times the compute of a normal reply.
Sources: OpenAI: Reasoning models, Anthropic: Thinking, Google: Gemini thinking, DeepSeek-R1 paper (2025), NVIDIA: How scaling laws drive smarter AI