Reasoning Models

Reasoning models are LLMs post-trained with reinforcement learning on verifiable rewards, producing extended internal "thinking" tokens that systematically improve performance on math, code, and planning benchmarks.

Why It Matters

Reasoning shifts the bottleneck from training compute to inference compute. Pricing, latency, and product UX all rebuild around the new dimension of "how long should the model think?"

Timeline

1. 2024-09

   OpenAI o1-preview launches; first commercial reasoning model.

2. 2025-01

   DeepSeek R1 open-sources reasoning at frontier quality, collapsing the moat narrative.

3. 2025-03

   Reasoning becomes default mode in flagship APIs (o3, Claude 3.7 Sonnet thinking, Gemini 2.5).

4. 2025-09

   Per-task reasoning budgets appear as first-class API parameters.

• 6 monthshigh confidence

  Every major frontier API exposes a "reasoning_effort" parameter as standard.

• 12 monthsmedium confidence

  Reasoning-token cost drops 10x year-over-year.

Related Trends

• Agentic Workflows

  Rising

  LLMs that plan, call tools, and complete multi-step tasks autonomously.

• Inference Cost Collapse

  Mainstream

  Token prices falling 80–95% per year while quality keeps rising.

• Open-Weights Frontier

  Rising

  Open-weight models closing the gap with closed frontier labs to under 6 months.