Chain-of-Thought
Also known as: chain-of-thought prompting, CoT, reasoning.
In plain English
LLMs generate one token at a time, so forcing them to lay out intermediate steps gives them more room to compute — each written step becomes context the model can build on, instead of leaping to an answer in one shot. Early on this was a prompt trick ('let's think step by step'). By 2026 it's built into a class of reasoning models (OpenAI's o-series, DeepSeek-R1, Claude's extended thinking) that automatically produce a long internal chain of thought before answering, trading extra time and tokens for markedly better results on hard problems. The reasoning is often hidden from the user by default. The trade-off is cost and latency: chain-of-thought can multiply token usage several times over, so it's worth it for genuinely hard tasks and wasteful for simple ones — which is exactly the kind of decision a good router makes automatically.
Example
Asked 'A shirt costs $40 after a 20% discount — what was the original price?', a model answering in one shot might blurt '$48' (wrong). Prompted to reason, it writes: '$40 is 80% of the original. Original = 40 / 0.8 = $50.' The visible working exposes and corrects the shortcut. On multi-step word problems this step-by-step approach can lift accuracy from roughly 50% to over 90%.
Chain-of-Thought in Vincony
Vincony's catalog includes the leading reasoning models that use chain-of-thought automatically — GPT-5.2, Claude Opus 4.5, DeepSeek-R1 — and Smart Routing reserves them for the hard prompts that justify the extra thinking time, sending routine work to faster, cheaper models.
Browse reasoning modelsTry it — 750+ distinct models across 80+ providers on one account
Vincony bundles GPT-5, Claude, Gemini, Perplexity Sonar Pro, DeepSeek, Mistral, and 750+ other models on one $0/month account.