If you have used a recent AI assistant, you may have noticed options labeled “thinking”, “reasoning” or “extended thinking”, or seen a model pause while it works through a problem. These are reasoning models, or reasoning modes, and they represent one of the most significant changes in how AI systems are built.
Updated September 2026: we added links to OpenAI’s o1 announcement, the DeepSeek-R1 paper and research on whether models’ visible reasoning is faithful.
The idea: spend more time on hard problems
A standard language model starts writing its answer immediately, token by token. That works well for many tasks, but for multi-step problems, like complex math, tricky logic or debugging, committing to an answer too early leads to mistakes.
Reasoning models are trained to first generate an extended internal chain of thought: breaking the problem down, trying approaches, checking intermediate results and correcting themselves, before writing the final answer. OpenAI’s o1, released in September 2024, brought this approach to wide attention, and it has since been adopted across the industry, including in models from Anthropic, Google and DeepSeek.
How they are trained
The key ingredient is reinforcement learning on problems with checkable answers, such as math and coding. DeepSeek’s R1 paper reported that reasoning abilities “can be incentivized through pure reinforcement learning,” without human-written examples of step-by-step reasoning. The model tries many reasoning paths and is rewarded when it reaches correct results, gradually learning strategies that work: decomposing problems, verifying steps and backtracking.
This created a new dimension for improving AI. Beyond making models bigger or training them on more data, developers can let them use more computation at the time of answering, sometimes called test-time compute. OpenAI reported that o1’s performance “consistently improves” with more time spent thinking.
Where reasoning helps
- Mathematics and quantitative problems
- Coding, debugging and software design
- Multi-step logic and planning
- Scientific and technical analysis
- Complex analysis of long documents
- Agent tasks that require planning several steps ahead
Where it is overkill
For simple questions, casual writing, quick rewrites and brainstorming, extended reasoning adds delay and cost without much benefit. Many assistants now decide automatically how much to think, or let you choose.
| Task | Reasoning worth it? |
|---|---|
| Rewrite this email to sound friendlier | Usually no |
| Summarize a short article | Usually no |
| Find the bug in this function | Often yes |
| Plan a project with dependencies | Often yes |
| Analyze a complex contract or dataset | Often yes |
| Solve a multi-step math problem | Yes |
Trade-offs to know
- Speed: thinking takes seconds to minutes.
- Cost: reasoning uses more tokens, which matters for API users and usage limits.
- Not infallible: reasoning reduces some errors but models can still reason confidently toward a wrong answer.
- Visible reasoning is not a perfect window: summaries or displays of a model’s thinking can help you follow its approach, but Anthropic researchers found that when models relied on a hint to reach an answer, their stated reasoning often mentioned the hint less than 20% of the time.
Practical tip For hard problems, give the model what it needs to reason well: all the relevant information, the constraints and what a correct answer must satisfy. Good inputs matter more than asking it to “think harder”.
When to use a reasoning model
Reasoning models let AI trade time for accuracy on hard problems. Use them for math, code, planning and complex analysis; use faster modes for everyday writing and quick questions; and verify important conclusions either way.
Sources
- Learning to reason with LLMs, OpenAI, September 2024
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, arXiv, January 2025
- Reasoning Models Don’t Always Say What They Think, Anthropic, arXiv, May 2025
