A reasoning model is a large language model that has been specifically trained — and given extra inference-time compute — to produce a long internal chain of intermediate reasoning steps before it commits to a final answer. Where a standard chat model tries to produce a good response in one pass, a reasoning model effectively “shows its work” first: breaking a problem into sub-steps, checking itself, and backtracking when a step looks wrong, all before the visible answer appears. This is often called test-time compute or inference-time scaling, since the model is spending more compute at the moment it answers rather than only during training.

The tradeoff is straightforward: reasoning models are slower and more expensive per query than a standard model of similar size, but they meaningfully outperform standard models on math, coding, multi-step planning, and anything requiring careful logical chains — the kind of work a founder or engineer would actually delegate to AI rather than a quick factual lookup. OpenAI’s o-series, Anthropic’s extended-thinking modes, Google’s deep-think modes, and open-weight models like DeepSeek-R1 all popularized this category starting in 2024-2025.

For a builder deciding which model to call for a given task, the practical question is usually “does this task need reasoning, or just retrieval and formatting?” — using a reasoning model for a simple lookup wastes time and money, while using a standard model for a genuinely multi-step problem risks a shallow, wrong answer.

🇵🇭 Philippine Example

For Philippine founders, reasoning models are most useful for tasks where a wrong step compounds — like calculating BIR tax withholding across multiple income brackets, structuring a cap table across several funding rounds, or planning a multi-stop delivery route — situations where the extra latency and cost of "thinking mode" is worth paying for correctness.

Related Terms