Perplexity AI Updates: November 5, 2026

1. Perplexity’s Decisions API Can Now Search the Web Before It Answers

Perplexity added a Web Grounding mode to its Decisions API, a decision-model API that returns a probability for each question asked about the state a caller sends (text, JSON or images) instead of generated text: the yes probability for a noul (yes or no) question, a probability per option for choice, and a probability per rubric level plus an expected score for score. Adding "mode": "web_grounding" to a POST https://api.perplexity.ai/v1/decisions request with the pplx-decider-v1.2 model makes Perplexity Fast Search run one query per question before the model call, merge and deduplicate the results, and append each page’s title, URL and text to the state as evidence; the response lists what the model read in search_results, a URL passed as the state is used as the evidence itself, and if a search fails or times out the model answers from whatever evidence came back. Each completed search costs $0.001 plus input tokens (search evidence counts as input), and a grounded request can carry up to 100 questions. On 328 true-or-false claims about October 2026 news and rare Wikipedia facts, each paired with a false twin, Perplexity reports accuracy rising from 56.4% without grounding to 97.3% with it (errors falling from 143 to 9), at the cost of median latency rising from 139 ms to 366 ms and p95 from 514 ms to 742 ms. On the 100 public questions of the PWS-Decisions benchmark, the grounded API answered 86 correctly against 51% without grounding, with a 370 ms median and 685 ms p95 measured from Perplexity’s own client. Source