Reasoning Models
A new class of AI models that think before answering, trading speed for accuracy on hard problems
The Rise of Reasoning Models
In September 2024, OpenAI released o1, the first widely available reasoning model. Unlike standard LLMs that generate responses in a single pass, reasoning models produce extended chains of thought — exploring approaches, checking work, and backtracking when needed. This represented a fundamental shift in where compute is invested: from pre-training (bigger models, more data) to inference time (more thinking per question).
The trend accelerated rapidly. DeepSeek-R1 demonstrated that open-source reasoning models could match proprietary ones. Anthropic integrated extended thinking into Claude. Google added thinking capabilities to Gemini. By early 2026, reasoning models were the default for complex tasks across coding, mathematics, science, and planning.
Why This Matters
Reasoning models opened a second scaling dimension. When pre-training scaling began hitting diminishing returns, inference-time compute provided an alternative path to better performance. The practical impact is significant: problems that stump standard models — multi-step coding challenges, graduate-level science, complex planning — become tractable with reasoning models.
The tradeoff is cost and latency. Reasoning models consume more tokens and take longer to respond. This creates a natural division: standard models for fast, simple tasks; reasoning models for hard problems where accuracy justifies the compute.
Related Reading
- Reasoning Models (Glossary) - Technical definition and how they work
- Scaling Laws - The pre-training scaling that reasoning models complement