Perplexity AI is releasing WANDR, a benchmark designed to evaluate how well AI agents handle complex, multi-layered research tasks. The open-source release gives developers and researchers a standardized way to measure whether their AI systems can actually do serious investigative work.
WANDR, which stands for Wide and Nuanced Deep Research, is built specifically for what Perplexity calls “wide research” tasks that require broad searching and deep investigation, aligned with Perplexity Computer’s capabilities.
What WANDR actually measures
The benchmark is designed to mirror the demanding workloads typically found in professional research environments, going well beyond simple question-answering or single-document summarization.
Perplexity positions WANDR as an evolution of earlier evaluation frameworks like WideSearch, reflecting a shift towards more realistic professional workloads.
The benchmark arrives alongside Perplexity’s “Search as Code” framework, or SaC, which treats research queries as programmable workflows rather than simple search strings. In preliminary evaluations on June 1, 2026, Perplexity’s SaC implementation scored 0.386 on the WANDR benchmark. The next-best score was 0.152, meaning Perplexity’s system outperformed its nearest competitor by a factor of roughly 2.5x.
The Perplexity Computer connection
WANDR is directly tied to Perplexity Computer, the company’s AI agent product that orchestrates multiple models to handle complex research and workflow tasks. The benchmark serves as the evaluation layer for the kind of work Perplexity Computer is designed to do.
Back in February 2026, Perplexity open-sourced the DRACO benchmark, another evaluation tool aimed at advancing collective understanding of AI capabilities. WANDR follows that same pattern.
As of July 11, 2026, xAI’s Grok 4.5 achieved top scores on WANDR when used in conjunction with Perplexity Computer, demonstrating that third-party systems can perform well on the benchmark.