Perplexity released WANDR, an open benchmark and evaluation harness built from 500 data-collection tasks meant to test whether a research agent can search broadly and dig deep without its answers degrading from one record to the next. The tasks mirror work knowledge teams already assign agents: competitor mapping, due diligence, talent sourcing, product comparisons, each one requiring dozens to thousands of separately checkable records. Perplexity’s own Search as Code system topped the leaderboard the company built to score them.
The distinction Perplexity draws is between wide research, discovering a large and often open-ended set of qualifying entities, and deep research, backing every one of those entities with specific, checkable evidence. A task can ask an agent to find 70 companies with a named executive appointment, then attach a sourced page proving the appointment and a separate page proving the company’s listing status. Every branch of that structure gets graded independently against the page and excerpt the agent actually cites, not against a fixed answer key.
The results argue that wide-and-deep research remains largely unsolved. Search as Code reached 0.363 soft F1 and 0.133 hard F1, the best score in the six-system comparison. Anthropic’s system placed second at 0.249 soft F1 and 0.072 hard F1. Every other system tested, including OpenAI’s and Exa’s offerings, topped out at 0.121 soft F1 and 0.035 hard F1. Perplexity’s own numbers mean its leading agent earned full credit for roughly one in seven of the records it submitted and one in seven the task actually required.
Perplexity is grading its own product here, and its search-orchestration system happened to win the benchmark it designed, wrote the grader for, and published first. Anthropic’s system came closer on raw quality while spending far more time, money, and tokens per task, a tradeoff Perplexity’s writeup acknowledges rather than hides. That transparency is worth noting. It does not change the fact that the only independently repeatable claim here is Perplexity’s own account of a contest it ran.
Scale is where every system breaks down fastest. Perplexity’s hard precision fell from 0.235 on the smallest task bin to 0.096 on the largest; its hard recall fell from 0.219 to 0.079. Adding intermediate steps to the task hierarchy hurt more: going from no intermediate key to three or more, Perplexity’s hard precision dropped from 0.392 to 0.019. Raising the effort setting helped some systems and not others. Perplexity’s soft F1 climbed to 0.447 at its highest effort tier, while cost across all systems spanned from $0.03 to $324.83 per task depending on configuration.
Any team currently buying a deep-research or agent API to build large, evidence-backed lists, not single-answer lookups, should run its own candidate vendors against WANDR’s published tasks rather than take a single vendor’s leaderboard position at face value. The benchmark’s real value is not the ranking; it is the diagnostic breakdown showing whether a given failure comes from insufficient discovery, weak enrichment, or excerpts that do not actually support the claim attached to them.
Perplexity published the WANDR benchmark and its findings on its research site on July 14, 2026.
