Don't add chain-of-thought to reasoning models — give only the end goal

Wharton Generative AI Labs·high signal

On models with built-in reasoning (o3-mini, o4-mini class), explicit chain-of-thought prompting produced only ~2.9–3.1% average accuracy gains while adding 20–80% more latency and cost — a net loss. The 2026 rule is to match prompting strategy to model class: reserve CoT for non-reasoning models (where it lifts Gemini Flash 2.0 ~13.5% and Sonnet 3.5 ~11.7%), and for reasoning models supply only the objective and constraints, letting the model allocate its own thinking budget.