Perplexity has rolled out its Decisions API and open-sourced pplx-decider-v1-27b, a multimodal decision model designed to return probability distributions over developer-defined answers rather than generate standard text.
The model is positioned for tasks such as classification, request routing, and selecting the next action inside an agent workflow.
Pricing and use case
Perplexity said the API costs $0.04 per 1 million input tokens, while output is free. The company described this class of model as a tool for handling frequent small decisions inside agent systems, including assigning a customer support ticket, deciding whether a tool should be called, and choosing the next branch in a workflow.
That differs from a traditional large language model, which typically generates a block of text first and then requires the system to parse an answer from it. A decision model returns preset options and their probabilities directly.
Benchmark results published by Perplexity
In Perplexity’s published evaluation, which covered 11 tests and 7,210 samples, pplx-decider-v1-27b reached an overall accuracy of 85.71%. That was above Jev at 84.51% and Qwen3.8-27B at 74.76%.
Perplexity also said the model did not lead in every category. Jev posted higher scores in 6 of the 11 tests, while pplx-decider-v1-27b came out ahead on FinancialPhraseBank, RAGTruth, TabFact, and Circa.
Not an independent benchmark
The published results came from Perplexity’s own testing rather than an independent third-party evaluation. The report also said AWS and Cloudflare have recently introduced similar models.