scraping-bench

scraping-bench grades web data APIs on whether an AI agent got the job done with them. Each provider becomes the only tool of the same agent (Claude Haiku 5.5, same prompt, same limits), which has to return specific facts: a page's first heading, a package version, every URL in a docs section. A script checks every answer against a key from an official source. No model grades anything.

This is campaign v2: four tracks (scrape, search, map and crawl) on the real web, mostly documentation sites, scored under the rules in docs/MEASURE-v2.md. Ten providers were tested: Browserbase, Context.dev, Firecrawl, Jina Reader, Linkup, Scrapfly, Spider, String, TinyFish and You.com.

Results site · Full report · How it's measured · Every row

On the site: scrape · search · map · crawl · every provider · head-to-head comparisons · Firecrawl alternatives · FAQ

Results

The agent score is the headline: the share of core tasks where every answer field was right. Intervals are 95% from 10,000 paired bootstrap draws that resample whole sites. A provider is in the top tier when its gap to the leader could be chance. "Answer in response" checks the provider's own output with no model involved, so it's an upper bound. Cost per finished task is all provider and model spend divided by tasks passed.

Scrape, 40 core tasks from 37 sites

Provider	Agent score	95% interval	Answer in response	Page coverage	Cost per finished task	Tier
Context.dev	38/40 (95.0%)	85.0% to 100.0%	38/40	98.5%	1.36¢	top
Spider	37/40 (92.5%)	81.0% to 100.0%	38/40	99.3%	0.88¢	top
Scrapfly	36/40 (90.0%)	78.6% to 100.0%	36/40	93.4%	1.17¢	top
String	35/40 (87.5%)	75.6% to 97.4%	38/40	98.2%	1.33¢	top
Linkup	34/40 (85.0%)	71.4% to 95.7%	37/40	99.6%	1.23¢	below
Firecrawl	32/40 (80.0%)	65.8% to 92.3%	36/40	97.8%	1.67¢	below
TinyFish	31/40 (77.5%)	61.0% to 92.1%	31/40	92.3%	0.71¢	below
Browserbase	30/40 (75.0%)	59.5% to 89.2%	30/40	82.0%	1.93¢	below
You.com	29/40 (72.5%)	56.1% to 89.2%	31/40	84.9%	1.00¢	below
Jina Reader	6/40 (15.0%)	5.1% to 25.6%	18/40	72.8%	4.26¢	below

Model alone, no tool: 0/40. Correct page given: 38/38.

Search, 30 core tasks from 18 sites

Provider	Agent score	95% interval	Answer in response	Cost per finished task	Tier
Linkup	28/30 (93.3%)	85.3% to 100.0%	23/27	2.19¢	top
TinyFish	26/30 (86.7%)	70.0% to 100.0%	25/27	0.97¢	top
Firecrawl	26/30 (86.7%)	71.4% to 97.1%	22/27	2.91¢	top
Context.dev	25/30 (83.3%)	62.5% to 100.0%	23/27	1.39¢	top
String	23/30 (76.7%)	55.9% to 96.2%	19/27	1.32¢	below
You.com	21/30 (70.0%)	50.0% to 86.7%	16/27	2.64¢	below
Browserbase	21/30 (70.0%)	50.0% to 86.7%	11/27	2.79¢	below

Model alone, no tool: 5/30.

Map, 15 core tasks from 15 sites

Provider	Agent score	95% interval	Answer in response	Cost per finished task	Tier
String	14/15 (93.3%)	80.0% to 100.0%	14/15	0.98¢	top
Firecrawl	13/15 (86.7%)	66.7% to 100.0%	13/15	1.32¢	top
Context.dev	12/15 (80.0%)	60.0% to 100.0%	12/15	0.98¢	top

Model alone, no tool: 0/15.

Crawl, 12 core tasks from 12 sites

Provider	Agent score	95% interval	Answer in response	Cost per finished task	Tier
Scrapfly	12/12 (100.0%)	100.0% to 100.0%	12/12	3.49¢	top
Firecrawl	12/12 (100.0%)	100.0% to 100.0%	12/12	8.29¢	top
Context.dev	11/12 (91.7%)	75.0% to 100.0%	11/12	5.86¢	top

Model alone, no tool: 0/12.

Spider ran scrape only. Map and crawl only include providers with a native endpoint for that job. Not tested in v2: Exa, Tavily, Bright Data, Perplexity and Parallel. Jina's search was left out because it has no hard cap on tokens per call.

String

String was added after the campaign. It ran on 2026-10-24, five days after the other providers, on the same frozen tasks and harness, and was measured in the same build as everyone else. No other provider was called again, and no other provider's number changed. String has no endpoint that returns page content across a site, so it isn't in crawl.

On scrape, String's own output held the answer on 38 of 40 tasks; the agent finished 35. The three tasks in between:

• pydantic-settings: the frozen key is 2.16.0, but 2.16.1 shipped on 2026-10-20. String returned the live page and the agent answered 2.16.1. Counted against the live key, String scores 36/40.
• Typer's contributing page and Context.dev's "Retrieve a batch" page: String's Markdown puts the HTML <title> as a # heading above the page's real H1, and the agent answered with the title.

results/measure-v2-string/drift.json re-checks every key String missed against its live source. The frozen-key scores above are the headline. results/measure-v2-string/reconciliation.json matches String's spend ($0.4341) and request counts to its dashboard.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗