Definition
A reasoning model is a large language model trained to work through a problem in intermediate steps before it gives a final answer, spending extra computation at inference time to plan, check and revise. OpenAI's API documentation describes them as models that 'generate an internal chain of thought to analyze the input prompt' and excel at complex tasks and multi-step planning, at the cost of being slower and more expensive than standard models (OpenAI).
The category became mainstream with OpenAI's o1 in September 2024, which was trained with reinforcement learning to produce long internal reasoning before responding. DeepSeek-R1, released with open weights in January 2025, showed that the approach could be reproduced outside the large US labs. Since then the line between 'reasoning' and 'regular' models has mostly disappeared. OpenAI folded its o-series into GPT-5 in August 2025, and by October 2026 the major frontier models expose reasoning as a setting rather than a separate product:
- OpenAI's GPT-6 family (GPT-6 Astra from September 3, 2026; Sol and Luna from September 22) runs in ChatGPT with a Thinking control that ranges from Instant to Extra High.
- Anthropic's Claude Opus 5.5, released September 22, 2026, has adaptive thinking that is always on, with an effort parameter that controls how deeply it thinks (Anthropic).
- Google's Gemini 4 Argon, released October 13, 2026, offers low, medium and high thinking levels, and Gemini 3.1 Deep Think remains Google's specialized mode for hard science and math problems.
Technically, the extra reasoning is a form of test-time compute. The model produces thinking tokens, sometimes visible as a summary and sometimes hidden, before writing the answer. Those tokens cost money (Google, for example, bills them as output tokens) and add latency, which is why products route simple questions to fast settings and reserve heavy reasoning for hard ones. The approach builds on chain-of-thought prompting, but the behavior is trained into the model rather than requested in the prompt.
For AI search, reasoning changes how answers are assembled. A reasoning model asked a research or buying question tends to plan several searches, read results, notice gaps and search again, which is the pattern behind query fan-out, deep research and agentic search. It is also more likely to cross-check a claim against several sources before repeating it. Anthropic reported that in an internal test where Opus 5.5 had to write a company earnings report from a copy of the web, 16 of its 18 reports passed a check where any invented figure or quote counted as a failure.
That has practical consequences for being cited. Pages that answer one specific sub-question clearly, with dates, figures and sources, fit the step-by-step retrieval pattern. Facts that differ between your site, review platforms and directories create conflicts a reasoning model may resolve against you or by omitting you. Comparisons with stated criteria hold up better than unsupported superlatives. When you track AI visibility, note which model and thinking setting produced an answer, because the same prompt can cite different sources at different effort levels.
Examples of Reasoning Models
- A procurement manager asks ChatGPT on a high thinking setting to compare three payroll providers for a 60-person company in Germany. The model runs several searches on pricing, local compliance and integrations before answering and cites nine sources.
- The same question on the Instant setting returns a shorter answer citing three sources, and one of the vendors that appeared in the high-effort answer is missing.
- A developer building a support bot on Claude Opus 5.5 lowers the effort parameter for simple password-reset questions to cut latency and token cost, and keeps higher effort for billing disputes.
- A SaaS company finds that a reasoning model describes its pricing inconsistently, traces the problem to an outdated plan table on a review site, and gets the listing corrected.
