fshy

fshy / Crawling

Jina Reader

jina.ai · Crawling · 1 record · JSON

Jina Reader Crawling

★ qualifiesanonymousno cardMper-ip-shaping

jina.ai/reader/ · reader and extractor API · crawler-scraper/jina-reader · JSON · compare

Free tierr.jina.ai is free at 20 requests a minute with no key; a free API key raises it to 500 RPM and carries 10M tokens
unit requests; metered on request-count
Pricenone stated
Per-object capnot applicable: the unit is a request, billed in output tokens once a key is used
Retention10M tokens per new key, one-time
Egressnot applicable: Markdown answers
Identityanonymous the keyless route needs nothing; a key is an account
Limits
  • request-rate 20 RPM keyless, 500 RPM with a free key, 5,000 RPM premium source
  • request-rate 10M free tokens per new key, counted on the output response source
  • per-object-cap a per-request token budget header bounds the answer; the OCR path costs 40 times more tokens source
  • per-ip-shaping 20 requests a minute per IP keyless, reported in x-ratelimit-limit as 20, 20;w=60; a domain can be blocked for a year or more after excess use source
Bindingper-ip-shaping
Interfacerest GET https://r.jina.ai/<url>; search at s.jina.ai; token budget per request
ToolingApache-2.0 source; Readability over linkedom; pdfjs-dist
Composes withRead step 2 in the broker's ladder; the only keyless renderer found
ReliabilityA domain can be blocked for years after excess use: nodeseek.com answered 403 with a block dated 2027-07-02. Page text arrives with the page's own instructions to models intact, so it must be treated as data.
NotesThe strongest keyless read leg found in this topic and the one that changes the escalation ladder: it read linux.do and a 2024 linux.do thread whole, keyless, where the broker's own read needed a provider and reported upstream_status 403. It is not universal. It refused www.nodeseek.com with a 403 AbuseAlleviationError stating the domain is blocked until 2027-07-02 after suspected excess use, so a per-domain block is the failure mode and it outlasts any retry budget. Every answer is a third party's rendering, so a number read from it is reported, not documented, until the origin confirms it.
Evidence rungmeasured
Checked2026-10-22 (live); stale on 2026-12-21
Parsed numbers
  • $.limits[3].value limit:per-ip-shaping 20.0 requests / minute
Evidence (4)
datesourcewhat it proved
2026-10-22jina.ai/reader/Reader API free; 20 RPM without an API key, 500 RPM with a free key, 5000 with premium; 10M free tokens per new key; s.jina.ai 100 RPM with a free key
2026-10-22r.jina.ai/https://example.com/200, 460 bytes, 0.9 s, keyless, Markdown with a title and a published time; the body carried two warnings about the page maybe not being fully loaded
2026-10-22r.jina.ai/https://linux.do/200, 30,172 bytes, 1,288 whitespace tokens, keyless, of a forum the broker's own read answers only through a provider with upstream_status 403; x-ratelimit-limit reads 20, 20;w=60 and x-usage-tokens 28 on a small page
2026-10-22r.jina.ai/https://linux.do/t/topic/49044200, a whole 2024 thread with every post body, 19 posts with author and date; the answer also carries the forum's own all-caps instruction block addressed to language models, so the page text is untrusted input