Reasoning models

A reasoning model generates internal deliberation before its visible answer, and how much it deliberates is a setting you control: effort on Claude, thinking_level on Gemini, reasoning tiers on GPT-5.6. This replaces the thing chain-of-thought prompting was invented to fake. Where you once had to trick a model into showing its working, you now turn a dial.

• Adding "think step by step". The reasoning already happened; the phrase mostly buys tokens and latency.
• Prescribing the method. Telling a reasoning model how to approach a problem constrains a process that is usually better than your description of it.
• Heavy scaffolding. OpenAI notes its current models need less of it, and that removing repetition improves results.
• Asking for the reasoning to be shown as a substitute for verifying it. Visible working is not the same as correct working.

• Describe the problem and the constraints precisely, then get out of the way.
• Set the reasoning depth with the provider’s own control rather than with prompt language.
• Say what a correct answer must satisfy. Criteria still help; instructions about method no longer do.
• Watch for over-thinking on easy tasks — the failure mode has flipped from too little reasoning to too much.

On small models, on open-weight models you host yourself, and on multi-step arithmetic or symbolic work, explicit step-by-step prompting still helps — and Meta lists it among its official Llama techniques for exactly that reason. The rule of thumb: if the model does not have a reasoning control, you are the reasoning control.