Perplexity has open-sourced its 27B parameter decision-making model, with official evaluations showing its overall score slightly outperforming Jev.

Beating AI Insight News: Perplexity has launched the Decisions API and open-sourced its multimodal decision-making model pplx-decider-v1-27b. Unlike standard large language models (LLMs), this model does not generate plain text; instead, it returns probability distributions directly from answers provided by developers, making it suitable for use cases like classification, request routing, and next-step action selection for AI agents. The API is priced at $0.04 per 1 million input tokens, with outputs offered free of charge. In 11 tests totaling 7,210 samples published by Perplexity, pplx-decider-v1-27b achieved an overall accuracy of 85.71%, outperforming Jev (84.51%) and Qwen 3.8-27B (74.76%). However, the model is not universally superior: Jev scored higher in 6 of the 11 tests, while Perplexity’s model led in benchmarks including FinancialPhraseBank, RAGTruth, TabFact, and Circa. Notably, these results come from Perplexity’s internal testing, not independent third-party evaluations. This category of models is tailored to handle high-frequency "micro-decisions" in AI agents, such as assigning customer support tickets, determining whether to call a tool, or selecting the next workflow step. Traditional LLMs typically generate full text before parsing answers, while decision models directly return predefined options and their associated probabilities. Similar models have also been launched recently by providers including AWS and Cloudflare.