Artificial Intelligence Made Simple

How to Prompt Reasoning Models Effectively

Recent research shows that traditional chain‑of‑thought prompting, once popular for base LLMs, often harms modern reasoning models. A Wharton study of 198 PhD‑level questions found CoT adds 20‑80% latency and can drop accuracy by up to 3.3% on Gemini Flash 2.5, while newer COLM 2025 research reports accuracy losses as high as 36.3% on pattern‑recognition tasks. Leading providers—including OpenAI, Anthropic, Google, and DeepSeek—explicitly warn against using CoT with reasoning‑oriented models. The article synthesizes interviews and research to propose eight updated prompting rules that work across model families and promise higher‑quality answers with lower cost.