What happened
Perplexity says WANDR (Wide ANd Deep Research) contains 500 realistic data-collection tasks drawn from de-identified patterns in production usage. The tasks cover work such as competitive mapping, due diligence, literature review, market analysis, product comparison and talent sourcing. They range from dozens to thousands of independently verifiable records. Across the benchmark, the tasks require 170,495 source-backed records; the median task asks for 50 members, four records per member and 245 records overall.
The benchmark models tasks as qualification-key hierarchies. A task can require, for example, a set of companies, qualifying employees at each company, and URLs supporting each employee. A released example asks for at least 70 US-based companies whose CEO or CFO appointment was first announced between March 1 and April 30, 2026, with an authoritative appointment page for each and a separate recognized listing-authority page. That produces 140 required records. Each path ends in evidence that a grader can recheck.
Perplexity describes a four-stage construction pipeline: seeding reusable task patterns from de-identified requests, authoring specifications and judges through an author–critic loop, admitting tasks after checks for feasibility and judge quality, then curating a balanced set by topic, scale, topology and difficulty. WANDR grades claims against cited evidence rather than a fixed answer key. It refetches pages, checks whether claims are clear and in scope, whether excerpts appear on the page, and whether the page and excerpts support the requirements. Precision, quality-adjusted recall and F1 are reported at both soft and hard levels; hard scores require a member's full evidence subtree to be correct.
Perplexity evaluated six production systems on all 500 tasks under pinned configurations. Perplexity Search as Code led with 0.363 soft F1 and 0.133 hard F1; Anthropic followed at 0.249 and 0.072. The strongest other system reached 0.121 soft F1 and 0.035 hard F1. Search as Code averaged $5.20 per task and a 14.9-minute median solve time, with 3.82 million reported tokens per task. On a 45-task subset, its xhigh setting reached 0.447 soft F1 and 0.224 hard F1. Across effort settings, reported cost ranged from $0.03 per task for Exa low to $324.83 for Gemini max.
The analysis finds that recall trails precision for every system and that full-member scores fall sharply below partial-credit scores. For Search as Code, hard precision fell from 0.235 in the smallest target-volume bin to 0.096 in the largest; hard recall fell from 0.219 to 0.079. Perplexity also reports that 57.5% of its submitted excerpts failed to support everything their records claimed. The tasks, evaluation harness and technical report are available on GitHub.
What the archive adds
New Runtime's earlier coverage says Gemini 4 Carbon moved from testing and early access to public availability on November 10, 2026, and became available in Devin Desktop and CLI on November 11, 2026. Those notes concern a different story: the linked Perplexity announcement is WANDR, not a Gemini availability update. This is a new thread.
Trend signal
WANDR supports the view that agentic research quality must be judged by inspectable evidence and completed work, not by answer fluency alone. Its hierarchical record scores make missing discovery, incomplete branches and unsupported claims visible. This connects to New Runtime's pattern on scientific agents and artifact evidence, although WANDR evaluates knowledge-work research rather than scientific agents specifically.
Why it matters
For teams deploying research agents, a plausible narrative can hide an incomplete search or weak support for individual claims. WANDR's distinction between soft progress and hard completion makes that gap measurable: a system can collect some valid records yet fail to deliver a usable set at the requested scale. Its diagnostics also separate discovery shortfalls from evidence failures, helping teams decide whether to improve query planning, source qualification, identity handling or citation checks.
The benchmark is especially useful as a reminder to set acceptance criteria around the unit of work a team actually needs. If the deliverable is a market map or diligence table, measure coverage and support for every row, not just correctness on sampled answers. The reported costs and latency also make quality trade-offs concrete, though benchmark results should inform rather than substitute for testing on a team's own sources and requirements.
Limits and uncertainty
These are Perplexity's benchmark results and methodology; independent replication is not established by the announcement. The six systems were evaluated under pinned configurations, and the effort-setting comparison used a 45-task subset, so results may not generalize to other configurations or current product versions. Soft and hard scores depend on WANDR's task construction and grading rules. The announcement does not establish how the benchmark predicts performance in a particular organization's research workflow, or whether improvements on WANDR transfer to live work.
Skill to build now
Turn one recurring research request into a small evidence tree: define the entities to find, the qualifying fields required for each, and the authoritative source expected for every claim. Score a sample run twice—first for partial progress, then for fully supported records. That exposes whether the bottleneck is finding enough candidates or completing evidence for each one.
Reader takeaway
WANDR's strongest result is still modest on complete, evidence-backed research. For builders, the practical lesson is to evaluate agents at the level of the finished collection: did they find enough qualifying items, and can every required claim be checked against its cited source?