Reasoning models vs standard models
Ask a standard model how many trains can run on a single track section with a 4 minute signalling gap and a 90 second dwell, across three junctions with different speed limits, and it answers in two seconds with a confident number. It is often wrong. Ask a reasoning model the same thing and it goes quiet for twenty seconds, writes several thousand tokens of working you may never see, catches its own mistake halfway through, and then answers. It is usually right, and it cost you thirty times as much.
Since OpenAI's o1 in September 2024, followed by DeepSeek-R1, Claude's extended thinking and Gemini's thinking models, most providers offer both kinds. A standard model starts writing the answer straight away. A reasoning model first generates a long chain of intermediate reasoning, then the answer. Knowing which one a task needs is now an everyday engineering decision, and choosing the expensive one by default is one of the more costly mistakes a team can make.
Your support bot classifies 200,000 tickets a day into 40 categories, and a separate tool plans multi-step database migrations for engineers. Which of the two would you put on a reasoning model, and what would you measure to decide?
What "thinking" actually is#
A language model produces one token at a time, and each token gets roughly the same amount of computation. A hard problem, one that needs many steps where each depends on the last, does not fit in the computation behind a single token. Asking a model to "think step by step" helped because it let the model spend more tokens, and therefore more computation, before committing to an answer.
Reasoning models make that the default and train it properly. They are trained with reinforcement learning on problems with checkable answers, mostly maths, code and logic, where the reward is whether the final answer was correct. Over training, the model learns which ways of working lead to right answers: breaking the problem down, trying an approach, checking it, backtracking when it fails. The result is a model that spends tokens where they help.
That leads to the property that matters most: accuracy grows with thinking tokens. On hard maths and coding benchmarks, letting the model think longer keeps improving results well past the point where a standard model has given up, which is why this is often called test-time compute: spending more computation when answering, rather than only when training.