Jina Reader: what happened when we tested it

Partly works. With a free trial key, Jina Reader turned two static documentation pages into clean Markdown: no navigation leftovers, about 94-97% of sentences matched exactly (the rest were present with formatting differences), and code examples were complete. Three things a reader should know: (1) the advertised no-key tier refused our cloud computer outright, (2) the JavaScript test page came back empty on the default call and only worked when we bypassed the cache, and (3) PDF text kept hundreds of line-break hyphens. A fresh page took 5-8 s; repeats were under 2 s.

Fixed before the first call (plan saved at 10:43 UTC+8): read 4 public pages through r.jina.ai without a key and compare with ground truth we fetched ourselves. P1 Wikipedia 'Markdown' (MediaWiki API plain-text extract: 50 sentences of 40+ characters, 191 article links, 4 images), P2 Python json docs (42 '>>>' example lines), P3 quotes.toscrape.com/js/ (10 quotes that only appear after JavaScript runs; the static HTML has 0), P4 an arXiv PDF that Jina's own FAQ uses as an example. Pass marks: 95%+ sentence recall and at most 1 navigation leftover per page, 9 of 10 JS quotes, and a 22-request burst to check the 20 RPM limit.

curl "https://r.jina.ai/https://en.wikipedia.org/wiki/Markdown"
(also the Python docs, quotes.toscrape.com/js/ and arxiv.org/pdf/2310.19923v4 URLs; no headers, no key)

1. Run 1: Fetched the vendor page jina.ai/reader/ and saved the claims and the rate-limit table (10:43)
   Observed: HTTP 200. Claims and '20 RPM' without a key found in the page HTML. jina.ai/pricing/ returned 404. The words 'reputation' and 'anonymous' appear nowhere on the page.
2. Run 1: Wrote the test plan and pass marks, then fetched our own ground truth for the 4 pages (10:43-10:44)
   Observed: Wikipedia API, Python docs and quotes.toscrape.com all HTTP 200. The arXiv export API timed out after 30 s, so we used the arXiv abstract page (200) instead.
3. Run 1: Sent the 4 test URLs to r.jina.ai with plain GET requests, one after another (10:44:36-10:44:37)
   Observed: 4 of 4 returned HTTP 401 in 0.10-0.23 s, 373-byte JSON: 'You have been blocked from performing anonymous queries due to bad network reputation (AS54113). Please authenticate.'
4. Run 1: Waited 60 s, then retried with the vendor's own example URL (example.com), the POST form, and the faster 'X-Engine: direct' engine (10:45:58-10:46:00)
   Observed: 3 of 3 returned the same HTTP 401 message.
5. Run 1: Opened one test URL in headless Chrome to rule out a curl-specific block (10:46:20)
   Observed: HTTP 401, same message (screenshot).
6. Run 1: Stopped. Did not sign up for a key, change networks or use a proxy, and did not run the 22-request burst
   Observed: With 0 successful conversions there was nothing to compare and no rate limit left to measure.
7. Run 2: sent the same 4 pre-registered URLs with the free trial key (10:53:31-10:53:35)
   Observed: All HTTP 200 in 0.40-1.76 s. Tokens: Wikipedia 12,750, Python docs 9,342, JS quotes page 80, PDF 12,214. The JS page came back with only the header and footer and the warning 'This page maybe not yet fully loaded'. Such fast replies are probably Jina's cache (its FAQ says repeats within 5 minutes are cached).
8. Extra diagnostics (marked as extra, not part of the plan): JS page again with X-No-Cache, then with X-No-Cache plus X-Wait-For-Selector: .quote; Wikipedia with X-No-Cache for a fresh latency (10:53:54-10:54:14)
   Observed: JS page: all 10 quotes, 8.0 s and 7.0 s, 251 tokens each. Fresh Wikipedia: same content, 5.1 s, 12,536 tokens.
9. Compared every output with our ground truth using a script (published in the evidence)
   Observed: Numbers in the claim table below. We hand-checked every sentence the script marked as missing.
10. Modest 22-request check at the keyed limit: example.com, one request after another (10:56:02-10:56:07)
    Observed: 22 of 22 HTTP 200 in about 4.4 s, 341 tokens each. The x-ratelimit-limit header said 500 per 60 s; x-ratelimit-remaining read 499 on every reply.

• No-key access blocked by network reputation (HTTP 401, status 40103) from a cloud network, on the first request rather than after a rate limit; not mentioned on the product page.
• Default call on a JavaScript page returned only header/footer with the warning 'This page maybe not yet fully loaded' (probably a cached incomplete copy); X-No-Cache fixed it.
• Spaces around some links were dropped ('RFC[7763](…)–'), which breaks exact-phrase search.
• PDF output keeps line-break hyphens (421 'x-y' breaks in one 14-page paper).
• x-ratelimit-remaining stayed at 499 for 22 requests in a row, so the header is not a reliable counter.
• Wikipedia: about half of the article's linked pages (102 of 191 distinct) appear as links in the output; we did not check which ones were dropped.

• File: Test plan and pass marks, written before the first call
• Screenshot: Vendor rate-limit table: r.jina.ai 20 RPM without an API key
• Screenshot: HTTP 401 'bad network reputation' in headless Chrome
• Log: All r.jina.ai calls with time, HTTP code, size and duration
• File: Raw 401 response body
• File: Raw 401 response headers
• File: Vendor claims as captured, with exact page text
• File: Ground-truth numbers we collected for the planned comparison
• Log: Run 2 calls (key redacted), extra diagnostics and the 22-request check
• File: Run 2 metrics: sentence recall, boilerplate markers, links, images, quotes, PDF
• File: Comparison script used for the metrics
• File: Output excerpts (full outputs kept in the repo because of the source pages' licences)

• "Convert a URL to LLM-friendly input by prepending r.jina.ai." … "Boilerplate such as navigation, headers, footers, and ads is stripped, and the main content is converted to Markdown." (source)
• "The default engine renders the page in a headless browser so client-side JavaScript executes before extraction." (source)
• "The best part? It's free!" Rate-limit table: r.jina.ai without an API key, 20 RPM. (source)
• "Reader natively supports PDF reading. It's compatible with most PDFs … and it's lightning fast!" (source)

L0. Run curl "https://r.jina.ai/https://example.com". If you get HTTP 401 'bad network reputation' (common on cloud servers), get the free trial key from jina.ai/api-dashboard/reader and send it as an Authorization: Bearer header. For JavaScript-heavy pages add the header X-No-Cache: true.

Tool

Jina Reader · https://jina.ai/reader/

Vendor

Jina AI

Category

Research

Pricing

Freemium: No-key tier advertised at 20 requests per minute, but our cloud network was refused (HTTP 401). The founder got a free trial key from the API dashboard without signing up; with it every call worked and responses reported a 500-per-minute limit. Paid use is token top-ups (jina.ai/pricing/ returned HTTP 404, so no per-token price recorded). (checked 2026-10-20, pricing page)

Account used

Run 1: none. Run 2: free trial API key from jina.ai/api-dashboard/reader, issued to an anonymous visitor without signup (obtained by the founder; key never logged or published)

Access used in test

L0 · No account access

Access in normal use

L0 · No account access

Tested by

Grok Bot (this assistant) on its own Linux computer with web search/fetch/shell/browser tools

Environment

Grok Bot on its own Linux computer (cloud box; outbound network AS54113 Fastly, Inc. per ipinfo.io), curl 8.14.1 and headless Chrome. Run 1 (10:43-10:47 UTC+8): no key. Run 2 (10:53-10:57 UTC+8): free trial key as an Authorization header, JSON responses so token use is reported.

Date tested

2026-10-20

Time

Run 1 about 4 min; run 2 about 4 min (10:53-10:57 UTC+8). Per call: 5.1 s fresh (Wikipedia), 7.0-8.0 s (JS page, fresh), 0.13-1.76 s when probably cached. The vendor's table says 7.9 s on average.

Cost

$0. Free trial key; 54,926 tokens used in run 2 (12,750 + 9,342 + 80 + 12,214 + 251 + 251 + 12,536 + 22 × 341).

Re-test by

2026-11-19