AI reasoning models can now show their uncertainty without slowing down or costing more

Researchers built a way to measure whether an AI reasoning model is confident in its answer by watching what it writes, not by running it multiple times or accessing its internal workings. This means anyone using a proprietary AI API can now know when to trust the model's reasoning and when to double-check it, without paying extra or waiting longer.

Until now, the only reliable ways to measure AI confidence required either running the same task ten times over (expensive and slow) or access to the model's internal machinery (which proprietary APIs don't give you). This method works on a single pass, watching for specific behavioral patterns in the reasoning trace itself. The practical effect is that production systems using closed-box AI reasoning models can now deploy confidence-gating at no additional cost. The paper shows the method catches high-confidence correct answers 96% of the time, and when combined with sampling for uncertain cases, achieves 90% accuracy at 71% coverage. This matters because it removes a major friction point in deploying reasoning models to real applications where knowing whether to trust the output is the difference between using the system and not.