Founding customers get 50% off the first year. Claim it →
← all posts

Jina Reader alternative for news articles

November 9, 20269 min readFeedmarkdown#jina-reader#markdown#llm#news-api#alternatives

If you already hold the article URLs, start with Jina Reader: one GET, and far cheaper per article than Feedmarkdown. If you need the URLs found, deduplicated and delivered, that is a different job. We ran 10 news articles through a feed on 2026-11-09, then 8 of them through Jina Reader on 2026-11-10 with a free API key. The two tools gave the same text on the clean pages. Each also failed in ways the other did not.

One disclosure first. We run Feedmarkdown. Its article text comes from WebCrawlerAPI's /markdown endpoint, which scrapes the page and has an LLM clean it. So this test mostly compares that endpoint with Jina Reader. For the HTML-to-Markdown token numbers behind both, see feeding news to an LLM.

If you want to follow a news topic and have no URLs yet, skip to where Feedmarkdown fits.

The results, article by article

Tokens are tiktoken 0.14.0, o200k_base. The table shows raw counts; "normalized" counts drop images and reduce links to anchor text. Boilerplate blocks (menus, cookie text, share buttons, related rails, ads, footers) were counted by hand. "Jina" is the keyed default output.

# Publisher Jina result Jina raw tokens Feedmarkdown result Feedmarkdown raw tokens Boilerplate blocks (Jina / Feedmarkdown)
1 valawyersweekly.com not sent - entity artifact, author-bio residue 1,072 - / 1
2 sg.news.yahoo.com full text, page edited since 999 clean 853 1 / 0
3 thegardenisland.com full page chrome 20,559 clean 1,636 89 / 0
4 greenvilleonline.com full article 1,716 full article, same sentences 1,228 1 / 0
5 valawyersweekly.com not sent - clean 1,513 - / 0
6 bostonglobe.com chrome, paywall prompt, then stopped 11,132 6 sentences missing, 1 entity artifact 752 133 / 0
7 bedfordgazette.com wrong page (a tracker) 157 2 paragraphs, then ROT47 text: unusable 5,910 0 / 0
8 a national news site 200, empty body, CAPTCHA warning 53 (headers only) author-bio and ad fragment 751 - / 2
9 chinadaily.com.cn same body text 559 clean 498 0 / 0
10 dailyherald.com same body text 1,043 entity artifact, author-bio residue 1,042 1 / 1

"Clean" is our label, not proof the text is unchanged.

The ten came from one Google News search (artificial intelligence when:1d, US English), run once: 20 found, 20 new, 0 deduped, 20 scraped, 0 errors. We took the first ten in API order, at most two per publisher, with the rule fixed before we looked at any output. They are conditioned on Feedmarkdown having scraped them, since a pull feed lists only successes; all 20 scraped.

Jina was attempted on 8, not 10. We did not send the two valawyersweekly.com articles (1 and 5) because that site's robots.txt came back as a Cloudflare 403 page we could not read. Of the 8, six returned a body from the right page (2, 3, 4, 6, 9, 10), one returned a tracker (7) and one returned nothing (8).

The Feedmarkdown text is from 2026-11-09. The Jina fetches ran 2026-11-10, 20:58 to 21:03 UTC, about 35 hours later. Article 2's page was edited in between (new headline, different Published Time), so that comparison is against a changed page.

What the numbers say, and how thin they are

Medians on 5 or 7 points are weak. On the 5 comparable pages (2, 3, 6, 9, 10), Jina's is 1,043 raw tokens (987 normalized) and Feedmarkdown's is 853 (619). Across the 7 Jina outputs with a body it is 1,043 vs 1,042.

What the table supports:

  • Clean pages match. On articles 9 and 10 the body text was the same in both, and normalized counts differed by 11% and 0.2% (as a share of Jina's count).
  • Chrome costs Jina on some pages. On articles 3 and 6, Jina kept the site's navigation, cookie text and footers: 89 and 133 boilerplate blocks. Feedmarkdown returned the article text.

Jina's tuned pass

On the keyed run we also did one tuned pass: X-Retain-Images: none, X-Retain-Links: text and X-No-Cache: true, per Jina's README. It removed link and image syntax: article 3 fell from 20,559 raw tokens to 3,279. Boilerplate tokens fell sharply too (18,666 to 1,429 on article 3, 9,999 to 1,851 on article 6). But the block count barely moved: 89 to 81, and 133 to 125. We did not use per-site selectors.

On article 4 the tuned response stopped at exactly 5,000 bytes, mid-sentence, and dropped 11 sentences the default pass kept. Our script set no byte cap, so this is the response as received. We did not retry it or find the cause.

Where Feedmarkdown got it wrong

  • Article 7 was mostly unreadable. Two readable paragraphs, then 26 blocks of ROT47-encoded text (about 7,900 characters), which is unusable and is why the count is 5,910. Jina returned a tracker page for the same URL, so neither tool gave a usable article.
  • HTML entities leaked into the text in 3 of 10 outputs. Article 10 has longing&#x73;*&#x20;*&#x66;or, which should read "longings for attention". They come from the HTML-to-Markdown step, which turns empty or space-edged <em> and <strong> tags into numeric references.
  • Bio or ad residue showed up in 3 of 10: author bios in articles 1 and 10, and a bio plus an ADVERTISEMENT](#after-bottom) fragment in article 8.
  • Article 6 lost text on both sides. Feedmarkdown omitted 6 sentences that Jina kept. Jina lost 16 that Feedmarkdown kept. All 6 were on the page, so Feedmarkdown's output was incomplete.

We could check Feedmarkdown's wording against the raw page on only 3 of 10 articles (3, 6, 9) and found no rewording. For the other 7 the page was blocked to a plain curl or never fetched, so "none found" does not mean "none".

Two real excerpts

Rule fixed beforehand: the first article where body text differs (now 6), and the first with at least 3 boilerplate blocks (3). Trimmed with [...].

Article 6, Jina keyed default (169 blocks including header lines; first lines and last two shown):

Title: As AI headwinds hit, Dartmouth leader remains a champion | Starting Point
Published Time: 2026-11-09T10:10:07.743Z
Markdown Content: / [Skip to main content](https://www.bostonglobe.com/2026/11/09/newsletters/starting-point/#main)
*    Sections / *           *   ### [Epaper](https://epaper.bostonglobe.com/?p1=BGMenu_Sections)
*           *   ### [Sports](https://www.bostonglobe.com/sports/?p1=BGMenu_Sections)
[... trimmed: nav, sign-in, headline, newsletter intro, paywall prompt, footer, cookie banner ...]
[![Image 33: Powered by Onetrust](https://cdn.cookielaw.org/logos/static/powered_by_logo.svg)](https://www.onetrust.com/products/cookie-consent/)
![Image 34: iiq_pixel](https://sync.intentiq.com/profiles_engine/ProfilesEngineServlet?at=20&mi=10&secure=1 [...]

Feedmarkdown, same page (17 blocks; first four):

# As AI headwinds hit, Dartmouth leader remains a champion
## The university's president is an oasis of optimism amid shaky markets, angry voters, and worried students
Dartmouth claims to have invented the term "artificial intelligence" in 1956.Cheryl Senter for the Boston Globe
### TODAY'S STARTING POINT
[...]

Jina kept the intro and bullets Feedmarkdown omitted, then hit the paywall prompt and lost the later sections.

Article 3, Jina keyed default (129 blocks including header lines; the first few and some footer lines shown):

Markdown Content: Opens in a new window Opens an external website Opens an external website in a new window
This website utilizes technologies such as cookies to enable essential site functionality [...]
Manage Preferences
Storage Preferences
Tuesday, November 10, 2026
[Today's Paper](https://printreplica.thegardenisland.com/)
[... blocks not shown: sign-in, menus, share buttons, article text, Trending, You Might Like, Around the Web ...]
[Terms of Service](https://www.carpentermediagroup.com/terms-of-service/)
2970 Haleko Road, Suite 205, Lihue, HI 96766. Telephone: (808) 245-3681

Feedmarkdown started at the headline and went straight to the article text.

The Jina Reader API call, limits and price

This is the keyed request we used; without a key, drop the header:

curl -sS -m 60 -H "Authorization: Bearer YOUR_JINA_API_KEY" "https://r.jina.ai/https://www.chinadaily.com.cn/a/202611/09/WS6af19b14e4b03ebf8519b587.html"

What we saw on 2026-11-10 for article 9: 200, 3010B, 1.396900s.

We used a free key, the "Toy Experiment" tier. What Jina's page (jina.ai/reader) said when we read it:

  • The rate limit is "20 RPM" without a key and "500 RPM" with a free key (read 2026-11-10). Our responses showed x-ratelimit-limit: 500.
  • Paid packages (read 2026-11-09): "1 Billion tokens" for "$50" ($0.050 per 1M tokens) and "11 Billion tokens" for "$500" ($0.045 per 1M tokens). The page does not use the words "minimum top-up". $50 is the smallest paid package shown.
  • The free "Toy Experiment" key card says "Non-commercial use only (CC-BY-NC)."

When to stay on Jina

Stay on Jina Reader if:

  • You already have the URLs.
  • You fetch them occasionally.
  • You want one GET per URL.
  • You need arbitrary sites. Feedmarkdown takes Google News and Google Alerts sources only. It cannot take your own URL list or a plain RSS feed today.
  • Price matters most. Jina's per-token price is far below ours per article:
Price Assumption Per article
Jina, smallest paid package $50 per 1B tokens 1,000 tokens per article $0.00005
Feedmarkdown Starter $19 per month 3,000 articles per month about $0.0063

On those assumptions, Feedmarkdown costs about 127 times more per article ($0.00633 / $0.00005). It is rough: Jina bills by its own count (its usage header said 21,481 for article 3 at keyed defaults), so chrome-heavy pages cost more. For a per-URL job, Jina wins on cost. For what a model then spends on those articles, see the cost to summarize 1,000 news articles.

What we did not test

  • A paid Jina key, ReaderLM, JSON output, s.jina.ai search, or per-site selectors.
  • The two valawyersweekly.com articles, which we never sent to Jina.
  • Rewording by Feedmarkdown's LLM cleaning on 7 of 10 articles.
  • Dedup in this run (0 merged), and the webhook.
  • Any other tool or day. It was one query, one machine in the Netherlands, one tokenizer and a two-per-publisher cap.

One person did the labelling, with imperfect blinding: link syntax and file size give the tool away. Boilerplate counts are a judgement call.

Where Feedmarkdown fits

A converter turns a URL into Markdown; it does not find URLs. Feedmarkdown's server handles the rest around the scrape:

  • Discovery. A Google News source returns redirect links. The adapter resolves them to publisher URLs before anything is scraped.
  • Dedup. Headline clustering runs on feed metadata before any paid scrape. Content similarity runs after. This run merged 0 stories (20 found, 20 new, 0 deduped), so it shows nothing about dedup; the evidence is in how often Google News repeats a story and deduplicating news by headline matching.
  • Delivery. An HMAC-signed webhook, or a JSON Feed 1.1, RSS or Atom feed you pull. This run used the pull JSON Feed.

The pull request and one trimmed item from this run (article 9), ids replaced:

curl -H "Authorization: Bearer YOUR_API_KEY" https://api.feedmarkdown.com/v1/feed/YOUR_FEED_ID/json
{
  "id": "YOUR_ARTICLE_ID",
  "url": "https://www.chinadaily.com.cn/a/202611/09/WS6af19b14e4b03ebf8519b587.html",
  "title": "AI models to improve earthquake risk assessment",
  "content_text": "# AI models to improve earthquake risk assessment\n\nChinese seismologists are developing [...]",
  "date_published": "2026-11-09T10:06:11Z",
  "date_modified": "2026-11-09T10:15:06Z",
  "tags": ["China Daily"],
  "_feedmarkdown": {
    "feed_id": "YOUR_FEED_ID",
    "source_id": "YOUR_SOURCE_ID",
    "source": "China Daily"
  }
}

The webhook sends a signed POST whose stable fields are title, url, source, author, published_at and markdown. We did not exercise the webhook.

It takes Google News searches and Google Alerts as sources only. It is paid, with a 7-day trial that needs no card and no free tier afterward. See the pricing section or start a 7-day trial.