Results at a glance
| Starting point | Details found | Runs with readable content | Median time | Time (P95) | Confirmed API charge | Charge per 100 details found |
|---|---|---|---|---|---|---|
| Exact page URL · Extract | 309/343 | 192/192 | 2.0 s | 8.4 s | $0.192 | $0.06214 |
| Company homepage · Search + Extract | 174/343 | 100/100 | 8.1 s | 17.7 s | $0.777 | $0.44655 |
| What we checked | Exact page URL | Company homepage |
|---|---|---|
| Homepage titles | 76/82 | 75/82 |
| Product-page titles | 72/75 | 15/75 |
| Prices with billing units | 56/59 | 28/59 |
| Customer quotes | 105/127 | 56/127 |
We matched returned text against 343 expected details from 82 companies. A run counted as readable when the scorer found nonempty text on an eligible page. Navigation can qualify as readable text even when a page's main content is missing. The charge per 100 details uses confirmed charges divided by matches.
Test 1: Reading a page you already have
Result. Parallel matched all seven expected prices on Momentic’s pricing page in 2.0 seconds for $0.001. For the supplied Vapi Rest customer story URL, Parallel returned the page title, navigation, and conference promotions. The story body was absent, and none of its 11 expected quotes matched. That case took 5.0 seconds for $0.001. Momentic pricing · Vapi customer story.
All 34 misses had readable text from their reference page. The expected details did not match that text under the benchmark rules.
Setup. We passed one known page URL to parallel.web.extract through treg and requested full content. We scored the returned full-content field.
Test 2: Starting from the company homepage
Result. Starting from Momentic’s homepage, the search surfaced its pricing URL. Extract returned nine pages and matched all seven prices in 3.4 seconds for $0.01. The Roo customer story was absent from those results, and all three expected quotes missed. Momentic pricing · Roo customer story.
Of the 169 missed checks, 157 had no readable text from their reference page in the response. The other 12 missed despite a readable reference page.
Setup. We ran parallel.web.search in fast mode using the homepage's host as the query seed and limiting results to the same site. We sent the homepage and up to nine distinct URLs from those results to parallel.web.extract for full content. The page cap applies to the URLs selected for extraction. The 90-second window covers both steps. All 100 homepage cases contribute to the runtime and charges.
When to use Parallel for this task
Use Extract when you already have the official page URL. When starting from a homepage, check the returned URLs for the specific pages that support the fields you need. On Momentic, searching the domain surfaced pricing and missed the Roo story.
Data and scoring
The November 17 tests through treg reuse the frozen November 15 URLs and reference checks. A hit is a normalized text match on an eligible URL. The pricing checks require the amount and billing unit. The November 15 Firecrawl and Context.dev tests ran through AgentRouter. The timings and charges here describe the November 17 workflows through treg, including their cache settings, concurrency, and retry handling.
Individual runs and scorecard lists inputs, returned URLs, checks, timings, and fees. Filter for provider parallel. Original comparison and methodology explains the shared dataset and scoring.