OpenAI’s GPT-5 Has a Reasoning Mode That Pauses to Ask You Questions

The most quietly significant feature in GPT-5’s launch isn’t the performance improvements. It’s the new clarification behaviour.

In clarification mode, GPT-5 pauses before executing long-horizon tasks and asks targeted questions about ambiguous aspects of the request. Not generic questions. Specific questions about the parts of the instruction where proceeding without confirmation is most likely to produce a result the user didn’t want.

This is a meaningful departure from the “attempt everything, correct later” pattern that has characterised LLM assistants since GPT-3. In agentic contexts, where a single model call triggers dozens of downstream actions, proceeding confidently on a misunderstood instruction is disproportionately costly.

The implementation is not perfect. In our testing, GPT-5 occasionally asked for clarification on instructions that were clear, and occasionally didn’t ask when it should have. But the direction is correct.