What is Reasoning Model?
Also called thinking model, reasoning LLM.
A reasoning model is a language model trained to spend additional computation on internal deliberation before answering, typically producing extended intermediate reasoning that may be hidden from the caller. Several major providers offer such models alongside faster general-purpose ones, often with a setting controlling how much reasoning effort to spend. They cost more and respond more slowly.
The difference is in training and in inference budget rather than in a separate architecture. These models are trained, often with reinforcement learning on problems that have checkable answers, to work through a problem before committing. At inference they generate reasoning tokens that are billed and counted against the context budget even when the caller never sees them.
The gain is real but uneven. On mathematics, code, multi-constraint planning, and careful analysis, extra deliberation reduces errors substantially. On retrieval, formatting, summarizing, classification, and simple extraction, it adds latency and cost for little benefit. Well-built agents route by step: a fast model for routine passes and a reasoning model for the few decisions that are genuinely hard.
Two mistakes are common. The first is treating a reasoning model as strictly better, when for most agent steps it is slower and more expensive without improving the outcome. The second is trusting exposed reasoning summaries as an audit trail. They are generated text, sometimes a condensed rendering of hidden reasoning, and they do not prove how the answer was reached.
Reasoning models absorb into the model what chain of thought prompting previously asked for explicitly, which changes agent design. Prompts no longer need to instruct the model to think step by step, and long visible thoughts in the ReAct style become less necessary. What remains is the engineering question of which steps deserve the extra spend.
Key points
- Trained to deliberate before answering, often with hidden reasoning.
- Reasoning tokens are billed and consume context budget.
- Clear gains on math, code, and planning; little on simple tasks.
- Exposed reasoning summaries are not a reliable audit trail.
In practice
An agent handles two steps in one run. Deciding which of forty support tickets belong to the billing category is routine, so a fast model classifies them in one pass. Working out a refund amount across a partial month, a promotional credit, and a currency conversion is not, so that step is routed to a reasoning model whose extra deliberation earns its cost.