Perplexity releases Decider v1.3 models that return probabilities instead of text

Perplexity has released version 1.3 of Decider and Decider Flash, models that answer yes/no or multiple-choice questions with probabilities instead of writing text, . Developers use such scores to sort, filter, moderate or route requests, where a number and a threshold work better than a written answer.

The main addition is Locate, which returns the exact matching text and its position, for tasks such as finding names or personal information. The models are on Hugging Face, OpenRouter and Perplexity's Decisions API. The full model is built on a Qwen3.8-27B base, and Flash is a smaller version trained to copy it.

On a sample of one million real API requests, Perplexity says v1.3 agreed with frontier AI judges 93.98% of the time, slightly up from 93.68% for v1.2. Perplexity also uses Decider itself to review its agents' security logs.