What is a reasoning model?
A reasoning model is a language model trained, usually with reinforcement learning, to generate an internal chain of thought before answering, trading extra output tokens and latency for higher accuracy on multi-step problems.
OpenAI's "Learning to reason with LLMs" (September 2024), which introduced o1, described a model trained with large-scale reinforcement learning to use a chain of thought productively: breaking hard steps into simpler ones and catching its own mistakes. OpenAI reported that performance kept improving with more training compute and with more time spent thinking at inference.
DeepSeek-R1 (January 2025) showed the same approach in open weights, reporting that reasoning behaviour such as self-reflection and verification emerged from reinforcement learning without human-written reasoning examples. A month later, Anthropic's Claude 3.7 Sonnet put standard and extended thinking in one model. By late 2026, reasoning is less a separate model class than a setting on most frontier models.
Reasoning as a dial
Effort levels
OpenAI's GPT-6 Astra exposes effort levels from low to max, and xAI's Grok 4.7 from low to xhigh.
Adaptive thinking
Current Claude models decide per request whether and how much to think, steered by an effort setting. Anthropic's docs list Claude Opus 5.5 and Haiku 5.5 defaulting to medium effort.
Thinking by default
DeepSeek's API runs V4 Pro with thinking mode on by default.
Reasoning models inside agents
OpenAI's GPT-5 prompting guide (August 2025) uses reasoning effort to calibrate agentic eagerness, how much an agent explores before it acts, and recommends the Responses API so reasoning carries over between tool calls instead of being rebuilt each turn.
Anthropic's API requires the thinking block that accompanied a tool call to be returned, unmodified, with the tool result. Harnesses that rewrite or drop those blocks get errors. Interleaved thinking lets the model reason after each tool result, which helps most on long tool chains where the next step depends on what came back.
Where reasoning models fall short
Cost and latency
Thinking tokens are billed as output. On simple lookups or classifications, the extra spend buys little.
Fewer knobs
Several reasoning models no longer accept custom temperature or top_p values, so output is steered through prompts, effort and schemas instead.
Reasoning is not an audit trail
Anthropic's "Reasoning models don't always say what they think" (April 2025) found that Claude 3.7 Sonnet mentioned a hint it had used only 25% of the time, and DeepSeek R1 39% of the time.
How Motiontic's engineers apply it
Motiontic's engineers set effort per step from eval results: high where a step plans or makes a judgement call, low or off where it extracts or routes. The decision is logged with cost and latency per task, so it can be revisited when a model changes.
A reasoning model compared with a standard model call
| Standard call | Reasoning model | |
|---|---|---|
| How it answers | Writes the answer directly | Thinks first, then answers |
| Tokens per request | Lower | Higher, thinking is billed as output |
| Latency | Lower | Higher and more variable |
| Fits | Extraction, routing, short drafting | Planning, maths, code, multi-step judgement |
| Main control | Prompt and examples | Effort level, plus prompt |
| Typical failure | Skips a step it needed | Overthinks easy inputs, costs more than it returns |
Common questions
What is a reasoning model in AI?
A language model trained to produce a chain of thought before its final answer. It spends more tokens and time per request in exchange for better results on problems that need several steps.
Are reasoning models better for AI agents?
For planning and judgement steps, often. For simple, high-volume steps such as extraction or routing, a lower effort setting or a smaller model is usually cheaper with no loss in quality. Evals per step settle it.
Can the reasoning of a reasoning model be trusted?
Not as a full record. Anthropic's research found models often leave out factors that influenced their answer, so agents should be judged on their actions and outcomes, not on their stated reasoning.
What is reasoning effort?
A setting that controls how much a model thinks before answering. Higher effort usually improves accuracy on hard tasks and increases tokens and latency.
Related
Extended thinking
Claude reasoning in thinking blocks before it answers, now steered by effort.
Chain-of-thought prompting
Asking the model to show its working, now mostly built into the model itself.
Think tool
A tool that does nothing, so the model has somewhere to stop and think between actions.
Claude Opus vs Sonnet
Two Claude tiers compared on price, limits and Anthropic's own results, with a routing rule.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- OpenAI, Learning to reason with LLMs (2024-09-12)
- DeepSeek-AI, DeepSeek-R1 (arXiv 2501.12948, 2025-01-22)
- Anthropic, Claude 3.7 Sonnet and Claude Code (2025-02-24)
- Anthropic docs, Extended thinking
- OpenAI, GPT-5 prompting guide (2025-08-07)
- Anthropic, Reasoning models don't always say what they think (2025-04-03)
- OpenAI API changelog (GPT-6 Astra entry, 2026-09-03)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.