Chain of thought

A prompting technique and emergent capability of large language models in which the model produces intermediate reasoning steps before arriving at a final answer. Originally surfaced as an emergent capability in 2022; now a first-class feature of frontier models (Claude extended thinking, OpenAI o-series, Gemini Thinking, DeepSeek R1) where the model spends compute budget on reasoning tokens before responding.

Chain-of-thought is what the 2024-2026 reasoning-model wave operationalises. The capability matters for enterprise agent procurement because reasoning-model pricing is materially different from non-reasoning models — both because each call uses many more tokens and because the per-token rate is typically higher. The right procurement test: identify which workloads in the deployment actually need reasoning-mode (genuinely difficult judgement, multi-step planning, novel problem decomposition) and route those to the reasoning model; route everything else to the standard model. Routing-by-step rather than routing-by-deployment is the single biggest cost optimisation available in 2026 enterprise agent stacks.

Articles that analyse this term

• Production agentic AI cost: the layered optimisation playbook for enterprise CFOs
• GPT-5 Pro at $200 a month: what the pricing tier signals to enterprise IT

Primary sources

• Anthropic. Claude 3.7 Sonnet extended thinking
• OpenAI. Reasoning models — o-series