Reasoning Models and Extended Thinking
What changes when a model thinks before it answers: how reasoning and effort settings work, when they pay off, and how they change cost and latency.
Ask an older model a tricky logic puzzle and it blurts out an answer in the first sentence, often a wrong one. Ask a current model the same question and something different happens: it pauses, works through the problem privately, catches its own mistake, and then answers. That pause is thinking, and it's the biggest change in how models behave since chat assistants arrived.
Thinking makes models dramatically better at debugging, planning, math, and multi-step analysis, which is exactly the work agents do. It also costs money and time, and on simple tasks it's pure waste. In this lesson you'll learn what a reasoning model actually does, how effort settings control it, when extra thinking pays off and when it doesn't, how it changes your bill and your wait, and how to prompt a model that thinks for itself.
From answering to thinking first
In LLMs in Plain English you learned that a model writes one token at a time and can't go back to revise what it's written. For a hard problem, that's a handicap: the first words of the answer get committed before the model has worked anything out.
The fix people discovered early was to ask the model to "think step by step". Writing out intermediate steps gives the model scratch paper. Each step becomes context it can build on, so later steps are better informed.
A reasoning model builds this habit in. Before writing its visible answer, it generates a stretch of private working-out, usually called thinking or reasoning. In that space it breaks the problem down, tries approaches, checks its work, and backtracks when something doesn't add up. Then it writes the final answer, informed by all of that.
The important difference from just prompting "think step by step" is training. Reasoning models are trained with reinforcement learning on problems with checkable answers, such as math, code that must pass tests, and logic puzzles. They're rewarded when the final answer is right, so they learn what kinds of thinking actually lead to correct answers: checking intermediate results, considering edge cases, noticing contradictions.
A short history
- September 2024: OpenAI's o1 (first released as o1-preview) popularized the idea of a model that spends time thinking before it answers.
- Early 2025: in January, DeepSeek-R1 showed the technique in an open-weight model; in February, Anthropic added extended thinking to Claude with Claude 3.7 Sonnet.
- Today: thinking is standard in frontier models. Claude Opus 5.5 and Claude Sonnet 5.5 think by default, and most major model families offer a way to control how much.
So "reasoning model" is less a separate category now and more a capability that most serious models have, with a dial for how much to use.