Reasoning Models and Chain of Thought: AI That Thinks

Key Takeaways

What Changed with Reasoning Models

Standard LLMs generate text token by token without any explicit planning phase. They commit to their first token and build from there, which works brilliantly for most tasks but fails on problems requiring multi-step deduction, backtracking, or careful error correction. The model has no way to reconsider an early token once it's generated. This is why Chain of Thought prompting was developed as a workaround, forcing the Foundation Model to externalise intermediate steps before committing to an answer.

Reasoning models allocate additional compute at inference time by generating a 'thinking' token sequence before producing a final answer. During this thinking phase, the model explores approaches, catches errors in its own logic, tries alternative paths, and builds up to the answer incrementally. This is qualitatively different from simply generating a longer response. Reasoning models are also the backbone of Agent systems, where an AI must plan, use tools, and verify its own outputs across many steps.

How Reasoning Models Are Trained

Reasoning models are trained primarily through reinforcement learning on tasks with verifiable answers: competition mathematics, formal logic puzzles, algorithmic coding challenges. The model receives a signal only when it reaches the correct final answer, not for the quality of its intermediate reasoning. Over thousands of training steps, it discovers reasoning strategies that reliably produce correct answers.

This outcome-based training is key. Rather than being taught how to reason by human annotators (who might reinforce suboptimal reasoning patterns), the model discovers what works through trial and error at scale. The resulting reasoning chains often look different from human step-by-step working, they reflect what the model found effective, not what a human would explicitly write.

Key Reasoning Models

OpenAI o1 (September 2024) was the breakthrough: 83% on AIME math vs 13% for GPT-4. o3 (January 2025) improved further: 87.7% on GPQA Diamond (near human expert level), 96.7% on AIME 2025. o4-mini offers strong reasoning at a lower price point. Anthropic's Claude 3.7 Sonnet with extended thinking is competitive and uniquely shows users the thinking process.

DeepSeek R1 demonstrated that reasoning model training can be done outside OpenAI, matching o1 as an open-weight model. Google's Gemini 2.5 Pro includes deep thinking capabilities. Mistral's Magistral Medium is a European-developed reasoning model. The technique has spread rapidly across the field.

When to Use Reasoning Models

Use reasoning models for: hard mathematical problems, competitive programming, multi-step logical deduction, scientific analysis with complex inference, legal reasoning, architectural decisions that require exploring many options. The performance gap over standard models is largest for tasks where 'thinking more helps', problems where the answer isn't immediately obvious.

Don't use reasoning models for: simple Q&A, creative writing, summarization, straightforward coding, conversational applications. The added cost (both money and latency) isn't justified when standard models perform adequately. A practical routing strategy: try the cheapest capable model first, escalate to reasoning if the output fails quality checks.