AI Glossary

Reasoning models

LLM foundationsPublished By Simon Budziak

Reasoning models are large language models trained or prompted to generate an internal chain of intermediate reasoning steps before producing a final answer, trading extra inference time and cost for measurably better performance on math, coding, and multi-step planning tasks that trip up a standard model.

How is a reasoning model different from a standard LLM?

A standard LLM predicts its response token by token with no dedicated space to work through a problem first. A reasoning model is trained to spend extra inference budget generating intermediate steps before it commits to an answer, closer to working through a problem on scratch paper than answering off the top of its head. The category spans several implementations, OpenAI’s o-series and its reasoning modes, Anthropic’s Claude extended thinking, and open releases like DeepSeek-R1 all generate this kind of intermediate reasoning, though the exact mechanism differs by vendor and shifts often enough that the durable fact worth learning is the category, not which model currently leads it.

When is the extra latency and cost worth paying?

For tasks with a genuinely correct answer reached through several dependent steps, math, debugging, multi-step planning inside an agentic workflow, the accuracy gain is real and worth the wait. For a simple lookup, a format conversion, or anything latency-sensitive, a standard model at low temperature is faster and cheaper for the same or better result, which is why picking a temperature and a reasoning tier is a per-task decision, not a permanent default.

Frequently asked questions

Do reasoning models replace regular LLMs?

No. They trade speed and cost for accuracy on hard, multi-step problems, which makes them the wrong default for simple tasks a standard model already handles well and faster.

Which reasoning model should I use?

There is no single answer, since the category spans OpenAI's o-series, Claude's extended thinking, DeepSeek-R1, and others that shift often; pick based on the task's accuracy needs and your own evals, not brand loyalty to one vendor.

Summarize this page with
ChatGPTClaudePerplexity

See how this works in a real workflow