Perplexity Open-Sources WANDR — 500-Task Benchmark for Wide & Deep Research
Jul 14, 2026: Perplexity releases WANDR — 500 research tasks, 170,495 evidence records, Apache 2.0 on GitHub. SaC leads at 0.363 soft F1; hard F1 tops 0.133. explainx.ai maps grading, competitors, and Perplexity Computer fit.
On July 14, 2026, Perplexity (@perplexity_ai) open-sourced WANDR — Wide ANd Deep Research — the internal benchmark it built to train and measure Perplexity Computer's wide-and-deep research stack. The release ships 500 tasks, 170,495 required source-backed records, an Apache 2.0GitHub harness, and a blunt headline number: even Search as Code — Perplexity's best system — hits only 0.133 hard F1 on the full suite.
WANDR is the wide sibling of DRACO (deep long-form reports). Where DRACO asks "write a complete objective report," WANDR asks "find every qualifying company, employee, filing, or competitor — and prove each row with a page that actually says what you claim." explainx.ai maps task shape, grading, leaderboard reality, and why Perplexity published it now.
Wide + deep — the two capabilities Perplexity encodes
From Perplexity's X thread:
"Wide-and-deep research requires two capabilities. (1) Agents must search broadly enough to find all qualifying entities. (2) Agents must investigate deep enough to support every claim with evidence."
Mode
Failure mode WANDR catches
Wide
Stops at 5 good examples when the task asked for 70 companies
Deep
Finds entities but excerpts don't support appointment dates, listings, or roles
Both
Polished narrative built on incomplete coverage
Real jobs cited: competitive mapping, due diligence, literature review, market analysis, product comparison, talent sourcing — the work people already delegate to Computer + SaC, not chatbot trivia.
Hierarchical tasks — qualification key trees
WANDR represents requirements as hierarchical, independently verifiable records:
snippet
company(n) → employee(m) → url(k)
Pattern
Meaning
Flat list
n items, one evidence URL each
Nested search
Parent entities, each with child enrichment
Matrix
Repeated child labels under different parents
Multi-branch
Same company key ties appointment + listing subtasks
Example task ceo_cfo_appointments: Find 70 US companies with CEO/CFO appointments (Mar–Apr 2026) · 70 appointment URLs · 70 separate listing-authority URLs → 140 records, shared company key. Missing listing proof = incomplete member even if appointment page is perfect.
That structure is why hard F1 punishes agents harder than soft F1 — one broken branch zeros the whole member.
500 tasks — scale and difficulty
Stat
Value
Total tasks
500 public Harbor packages
Total records demanded
170,495 source-backed
Median breadth
50 members
Median depth
4.00 records per member
Difficulty split
167 / 166 / 167 (lower / middle / higher)
Tasks are seeded from de-identified production patterns — not synthetic Jeopardy prompts. Pipeline stages: seeding → authoring (author–critic loop) → admission (feasibility + judge audit) → curation.
Key insight from Perplexity: Difficulty is not scale alone — it depends on work per record (disambiguation, cross-source checks, regulatory nuance).
Reference-free grading — no stale gold answers
Fixed answer keys fail for open-ended research. WANDR instead:
Solve — normalize agent output
Fetch — retrieve each cited URL (browser retry on failure)
Judge — identity resolution + per-record verdict
Score — precision, recall, soft/hard F1 up the hierarchy
Each record: item + URL + excerpts + answer. Grader checks page usability, claim scope, excerpt fidelity, and whether evidence supports every requirement.
Diagnostic signal
What it means
High precision, low recall
Good rows found, not enough of them
Large soft → hard drop
Partial trees common; full completion rare
Retrieval-only vs full F1 gap
Finding a plausible page is easy; excerpt support is hard
Perplexity SaC: 0.531 retrieval-only soft F1 → 0.363 full soft F1. The gap is evidence construction, not search alone.
Leaderboard — SaC leads; nobody wins
Six production systems on all 500 tasks (pinned configs, same verifier):
System
Soft F1
Hard F1
~Cost/task
Median time
Perplexity Search as Code
0.363
0.133
$5.20
14.9 min
Anthropic
0.249
0.072
Higher $ + tokens
Slower
OpenAI
≤0.121
≤0.035
Faster, cheaper
Lower
Exa / Parallel / Gemini DR
≤0.121
≤0.035
Varies
Varies
Perplexity's June SaC article cited ~2.5× over next-best on WANDR — consistent with 0.363 vs ~0.15 class scores for others.
Unsaturated: Best hard precision ~0.150, hard recall ~0.134 — leader earns full credit for roughly one in seven requested members. Perplexity states plainly: "Wide-and-deep research is still a long way from solved."
Effort scaling (45-task subset): Perplexity xhigh reaches 0.447 soft / 0.224 hard F1; cost spans $0.03 (Exa low) to $324.83 (Gemini max) per task.
Four findings that matter for builders
1. Partial progress is common; complete coverage is not
Every system's soft recall sits below soft precision — denominators hurt once full volume counts.
2. Scale compounds failure
Largest volume bins: Perplexity hard precision falls 0.235 → 0.096; recall 0.219 → 0.079.
3. Deeper hierarchies are harsher
Three or more intermediate keys: Perplexity hard precision0.392 → 0.019.
4. Discovery bottlenecks before judging
Mean top-level discovery completion 0.611–0.951 across systems. Duplicate collapse loses only 0.017–0.205 points — under-delivery, not dedup, dominates.
Page vs excerpt failures: Only 3.2%–8.9% unusable pages (except OpenAI 23.1%). 57.5%–86.6% of excerpts fail to support full claims — the hard part is evidence packaging, not link-finding.
Why open-source now — product + research loop
Motive
Detail
Legitimize Computer
Public harness for claims SaC already made in June research
Training factory
Pipeline generates sibling tasks · RL partial credit on branches
Competitive intel
Shows Anthropic closest on quality but costlier; OpenAI/Exa cheaper but weaker on coverage
Harbor ecosystem
Standard task packages · comparable to other agent benchmarks
Ties to July 9 GLM 5.2 orchestrator preview — WANDR is the scoreboard Perplexity used for cost vs Opus math on Computer tasks.
Overlapping "deep research" narrative — WANDR adds published harness + 170k record accounting
For teams building GEO / competitive intelligence pipelines, see GEO for marketers — WANDR is the eval layer for whether your agent actually covers the full competitor set, not just writes a convincing summary.
Summary
WANDR (July 14, 2026) is Perplexity's open Apache 2.0 benchmark for wide-and-deep research agents: 500 tasks, 170,495 required evidence records, hierarchical qualification trees, and reference-free URL re-fetch grading. Search as Code leads at 0.363 soft F1 and 0.133 hard F1; Anthropic is second; every other tested stack scores lower. Hard F1 near 0.13 means the field — including Perplexity — is nowhere near solved.
For developers, the repo is a runnable spec of what Perplexity Computer optimizes for. For the market, it reframes "research agents" from one beautiful report to 170 thousand verifiable rows — a harder, more honest bar.
Scores, task counts, and system comparisons reflect Perplexity's July 14, 2026 release. Re-run benchmarks on pinned configs before citing leaderboard numbers in production decisions.
