Benchmark Almanac

Perplexity

We track 9 benchmarks and 1 arena that Perplexity made or helped make. Its models are on its lab page.

2 in use 3 with public code 6 with a leaderboard here

Benchmarks

Name Kind Released Best score Status
DRACO 100 deep-research tasks from real Perplexity queries across 10 domains, graded on about 40 expert rubric criteria each. Deep research Feb 2026 88.6% In use
WANDR 500 wide-search tasks: find tens to hundreds of entities and facts, then return structured records with citations. Deep research Jul 2026 82.1% In use
Answer Factuality (Perplexity) Models answer questions from search results. Graders check that answers use facts grounded in the results and resolve conflicting or missing information. Faithful to sources Feb 2025 85.1% Dropped
Readability (Perplexity) Models answer search questions. Graders rate whether answers are concise yet detailed and use markdown formatting well. Taste and style Feb 2025 85.9% Dropped
BrowseSafe++ Detectors judge 38,675 matched attack and control web pages built from real production prompt injections, in HTML, Markdown, accessibility-tree and screenshot views. Scored by F1. Jailbreaks and prompt injection Oct 2026 91.5% Not in cards
BrowseSafe-Adapt Browser agents face adaptive prompt injections across 1,024 tasks. The score is attack success within 64 attempts. Jailbreaks and prompt injection Oct 2026 30.7% Not in cards
BrowseSafe-Bench Detectors and browser agents face prompt injections hidden in realistic HTML pages, across 11 attack types, nine placements and three writing styles. 3,680 test samples. Jailbreaks and prompt injection Nov 2025 86.3% Not in cards
Q2D-Web Retrievers search 190 million web documents for queries that agents reformulate from real conversations. Questions that need search Sep 2026 79.3% Not in cards
Q2D-Chunk Retrievers find short passages that answer a query or supply the evidence needed to interpret the answer. Questions that need search Oct 2026 Not in cards

Arenas

Name Kind Released Best score Status
PPLX Online LLM Human Evaluation Models answer three sets of 50 prompts about helpfulness, factuality and freshness. Contractors compare responses and give a separate overall preference. Overall chat quality Nov 2023 Not in cards