Reasoning Model

Reasoning models (also called "thinking models") are LLMs trained to generate extended internal reasoning — a chain of thought — before producing a final answer. This dramatically improves accuracy on hard problems: math, coding, logic, scientific analysis.

How they differ from standard models: Standard LLM: prompt → response. Reasoning model: prompt → [internal thinking: tries approach A, finds flaw, tries approach B, verifies…] → final answer.

The thinking is often hidden from the user (Claude's extended thinking), sometimes visible (DeepSeek R1 shows its scratchpad), and always billed for at the output token rate.

Key models (mid-2026):

• OpenAI o3/o4-mini — flagship reasoning; best at math and coding
• Claude (extended thinking mode) — available on Sonnet and Opus models
• DeepSeek R1 — open-source, shows reasoning trace, rivals o1 on benchmarks
• Gemini 2.0 Flash Thinking — Google's efficient reasoning variant

When to use: Multi-step math, complex code planning, strategic analysis, hard logic puzzles. Not cost-effective for simple tasks — thinking tokens add significant cost.