Subindex vs Parallel.ai FindAll — Set-Finding Benchmark
Benchmark · Web set-finding · June 2026

One engine finds 3× more, in minutes instead of hours.

On a six-query set-finding benchmark, Subindex — our web row-discovery engine — recovers far more of the true answer set than every tier of Parallel.ai FindAll, while returning in real time and at a fraction of the cost. It wins on all three axes at once: recall, speed, and price.

3.1×
Higher recall
0.49 vs 0.16 entities found
11×
Faster
~13 min vs 2.4 hrs
17×
Cheaper
$1.74 vs $29.25
vs FindAll core — the tier Parallel recommends by default
The full comparison

Every FindAll tier, measured on the same queries

Recall is the share of approved ground-truth entities each system actually returned across the six queries — total entities found ÷ total answer set (268 entities). FindAll offers four generator tiers that trade cost for thoroughness; Subindex runs one configuration.

PipelineRecallCost · 6 queriesWall time
SubindexR15
0.49
$1.74 ~13 min
FindAllpreview
0.03
$0.60 18 min
FindAllbase
0.11
$6.51 64 min
FindAllcore
0.16
$29.25 2.4 hrs
FindAllpro · projected
~0.19
$240–660 3–9 hrs

Pro tier projected from Parallel's published pricing ($10 fixed + $1/match) and the recall-vs-tier trend; not directly run.

Efficiency frontier

More of the answer, for less money

Each dot is one pipeline. Right is better (more recall); lower is better (less cost). FindAll's tiers climb steeply in price while recall barely moves — Subindex sits alone in the bottom-right: the most complete answer at the lowest spend.

$1 $10 $100 Cost · log 0.1 0.2 0.3 0.4 0.5 Recall → cheap + complete preview base core pro · projected Subindex 0.49 · $1.74
Recall on a linear axis (0–0.55); cost on a log axis (six-query total).
Speed

Interactive, not overnight batch

The gap that matters most in practice: Subindex answers in about two minutes per query — fast enough to sit inside a live product. FindAll's higher tiers run for hours, built for offline jobs you kick off and walk away from.

Subindex~126 s/query
13 min
live
FindAll preview
18 min
FindAll base
64 min
FindAll core
2.4 hrs
FindAll proprojected
3–9 hrs
Methodology

Six hard set-finding queries, scored the same way

Set-finding — "list every entity that matches these criteria" — is the row-discovery step behind a research table. We chose long-tail queries where the answer set is scattered across the open web, not sitting on one clean list page. The number after each query is how many of its ground-truth entities Subindex returned.

  • Fortune 500 CEOs hired in 202520 / 96
  • Michelin 1-star restaurants, Manhattan 202562 / 76
  • OpenAI safety-team departures, 2024–202623 / 32
  • EU seed-stage AI-safety startups6 / 28
  • Upscale restaurants in Lawrence, NY11 / 21
  • S&P 500 top-10 movers, Q1 202610 / 15
Metric
Recall = approved ground-truth entities returned ÷ total approved GT, pooled across all six queries (268 entities). Every hit is an entity the system actually returned — nothing is credited that wasn't retrieved.
Systems
Subindex at its shipped R15 configuration, one run per query. FindAll run through Parallel.ai's public API across all four generator tiers; matched-status candidates counted directly.
Fairness
Same queries, same ground truth, same scorer for every system. Cost is each system's own billed total; wall time is measured end to end. Pro-tier figures are projected from published pricing, not run.
Date
Measured June 8, 2026. Ground truth was human-approved and frozen before the runs.
Full ground truth is public. Every one of the 268 answer-set entities, and exactly which systems found each, is listed at subindex.ai/s/findall_benchmark_GT.
Subindex · web set-finding benchmark 6 queries · 268 GT entities · scored June 2026