Perplexity has released WANDR, an open benchmark for testing how well AI research agents handle high-volume, evidence-heavy knowledge work. Published July 14, it measures the kind of jobs people already hand to research agents, competitive mapping, due diligence, literature review, and talent sourcing, where the task is to build a large collection and support every entry with specific proof.
Even at a high effort setting, the strongest system reaches just 0.363 soft F1 and 0.133 hard F1. In plain terms, the best research agent tested earns full credit for only about one in seven records it submits. Here is what WANDR measures, what it found, and why it matters for anyone relying on these tools.
What Does WANDR Really Test?
As you may know, most AI benchmarks test whether a model can answer one question well. WANDR tests something harder. It asks whether an agent can answer the same question hundreds of times without the quality collapsing.
The name stands for Wide ANd Deep Research. It is the companion to Perplexity’s earlier DRACO benchmark, which measured deep research, the ability to produce one accurate long-form report. WANDR measures the wide dimension. A task might require finding 70 companies, then a named executive for each, then an authoritative source page for each of those. Every complete path gets checked independently.
The benchmark is built from 500 realistic tasks drawn from actual product usage rather than synthetic prompts. The median task asks for 50 members and 245 total records. Across all 500 tasks, WANDR calls for over 170,000 source-backed records.
How Perplexity’s WANDR Grades AI Models
WANDR does not use a fixed answer key, which the team argues is a poor fit for research questions whose answers change over time. Instead, it grades each claim against the evidence the agent cites.
WANDR is constructed from de-identified production use cases covering people’s day-to-day research tasks like competitive research, due diligence, literature review, market analysis, product comparison, talent sourcing, and more. pic.twitter.com/y6dClkm2a6
— Perplexity (@perplexity_ai) July 14, 2026
For every record, the grader re-fetches the cited page and checks four things. Is the page usable? Is the claim in scope? Do the quoted excerpts actually appear on the page? And do the page and excerpts support every requirement? This makes failures traceable. A system can be diagnosed as failing at discovery, enrichment, page qualification, or evidence extraction, rather than just getting a single pass-or-fail score.
What WANDR Found & Why It Matters
Perplexity tested six production systems, including its own Search as Code, along with offerings from Anthropic, OpenAI, Google’s Gemini, and Exa. Its own system led at 0.363 soft F1. Anthropic’s Claude placed second at 0.249. Every other system topped out around 0.121.
Three findings stand out for anyone using these tools. First, partial progress is common but complete coverage is rare. Systems find some good records but rarely all the ones a task requires. Second, the problem compounds with scale. As target volumes and hierarchy depth grow, accuracy drops sharply, with the hardest nested tasks pushing even the leader’s scores near zero. Third, finding a usable page is easy, but turning it into complete evidence is hard.
For most systems, under 9 percent of pages were unusable, yet more than half of submitted excerpts failed to fully support the claims attached to them. There is no free lunch on cost either. Effort settings ranged from three cents per task at the cheap end to over 300 dollars per task for Gemini at maximum effort. No single system led on both quality and efficiency.
WANDR represents these requirements as hierarchical, independently verifiable records.
— Perplexity (@perplexity_ai) July 14, 2026
It consists of 500 research tasks that require 170,495 source-backed records across three tiers of difficulty. pic.twitter.com/OX5e1zuGbe
For anyone deploying AI research agents in real work (such as ChatGPT Work), WANDR is a useful reality check. These tools are increasingly marketed for exactly the tasks the benchmark covers, due diligence, market analysis, competitive research. The results suggest that for large, structured research jobs, current agents deliver partial answers that still need human verification, not finished work.
For the industry, the value is in the diagnosis rather than the leaderboard. By localizing where systems fail, whether in discovery or evidence construction, WANDR gives developers a specific target to improve against. The team also notes the benchmark could support reinforcement learning, using partial credit at each step to train agents that plan for coverage and catch their own gaps before stopping.
One honest caveat worth keeping in mind. WANDR is built and published by Perplexity, and Perplexity’s own system tops the results. That does not invalidate the benchmark, which is open and independently checkable on GitHub, but readers should weigh vendor-run benchmarks accordingly. The broader finding, that no system scores well, is the part that holds regardless of who leads.
You can also read – Anthropic’s Claude Study Shows AI Doesn’t Communicate the Same Way in Every Language
