Reasoning Model

What is a reasoning model?

A reasoning model is an LLM trained with reinforcement learning on verifiable problems (math, code, logic) to produce an extended internal “thinking” phase before its final answer. Introduced by OpenAI’s o1 (2024) and rapidly replicated (DeepSeek-R1, Claude’s extended thinking, Gemini thinking modes), the approach scaled a new axis: test-time compute — spending more tokens at inference buys more accuracy, not just bigger training runs.

What changed in practice

Reasoning models made agents dramatically more reliable: multi-step planning, error recovery, and long tool-use chains that previously needed elaborate scaffolding now work with a single capable model in a loop. Much of 2024’s framework complexity (multi-agent choreography, hand-built plan-then-execute pipelines) collapsed into “give a reasoning model good tools.”

When to use one — and when not

• Use reasoning for math, non-trivial code, debugging, planning, and anything multi-step where being wrong is expensive.
• Skip it for classification, extraction, summarization, and formatting — a standard model answers faster and cheaper at equal quality. Routing easy traffic away from reasoning models is a top-three cost lever.
• Most providers expose an effort/budget dial; treat it per-request, not globally.

Cost reality

Thinking tokens are billed as output tokens — the expensive kind — and a hard problem can burn thousands before the first visible word. The same query can cost 5–50× more with heavy reasoning, and time-to-first-token stretches from milliseconds to tens of seconds. Cap thinking budgets in production and monitor them like any other resource.

What people get wrong

• Using reasoning models everywhere. Paying for thinking on “extract the email address” is the most common silent budget leak.
• Prompting them like old models. Chain-of-thought incantations (“think step by step”) are redundant — the model already does it; clear task specs matter more.
• Trusting the visible reasoning as an explanation. Summarized thinking traces don’t necessarily reflect the actual computation; treat them as an aid, not proof.

Reasoning Models vs. Prompted Chain-of-Thought

The key distinction between a dedicated reasoning model and a standard model prompted to “think step by step” is that reasoning models are specifically trained — typically via reinforcement learning against verifiable reward signals, like whether a math answer is correct — to generate long, self-correcting chains of thought as a core learned behavior, rather than relying on prompting alone to elicit reasoning-like output. This training produces qualitatively different reasoning: dedicated reasoning models often generate reasoning traces that explore multiple approaches, catch and correct their own errors mid-chain, and allocate visibly more “thinking” to harder problems, behaviors that emerge from the training process itself rather than being explicitly programmed. The tradeoff is substantially higher latency and token cost per response, since these models may generate thousands of tokens of internal reasoning before producing a final answer — which is why reasoning models are typically reserved for tasks that genuinely benefit from deeper deliberation (complex math, coding, multi-step planning) rather than used as a default for every request.