Reader is the service behind the r.jina.ai prefix. You put it in front of a URL and get the page back as markdown, ready for a model. It is a popular way to feed a RAG index or an agent, and it is easy to start with. The questions that come later are how many pages a minute you can pull, what the bill looks like, and what you get from sites that do not want to be read.
Everything below comes from Jina's own Reader page, read in November 2026. We could not read its token price table, so this post gives no price per million tokens. Check the table before you budget.
Jina Reader API rate limits
| Endpoint | No key | Free key | Paid key | Premium key |
|---|---|---|---|---|
| Reader (r.jina.ai) | 20 per minute | 500 per minute | 500 per minute | 5,000 per minute |
| Search (s.jina.ai) | blocked | 100 per minute | 100 per minute | 1,000 per minute |
A limit trips when either the request count or the token count per minute is reached, and an IP based cap of 10,000 requests per 60 seconds also applies. The page lists an average Reader latency of 7.9 seconds, which matters more than it looks. At 500 requests a minute and 7.9 seconds each, you need about 66 requests in flight at once to use the limit.
How Jina Reader bills tokens
Reader counts the tokens in the response it sends back. A long article costs more than a short product page, and a page full of navigation text costs more than you expect. Failed requests are not charged. Search costs a fixed amount per request starting at 10,000 tokens, so the 10 million free tokens on a new key cover roughly 1,000 searches at that minimum.
Two settings change the count. OCR mode costs 40 times more tokens, and a token budget header caps the tokens per request, with a request that exceeds the cap failing. Neither is a problem on its own. They are the reason a Reader bill follows page length, not page count.
What happens on blocked pages
Jina's page says Reader does not actively circumvent website defense mechanisms, anti bot systems or access controls, and that a paid key does not grant access to additional websites. That is a clear position and an honest one. It also means a retail, travel or directory site that serves a challenge will hand your pipeline a challenge.
The cost of that is not the failed request, which is free. It is the challenge page that returns with a normal status and gets embedded as if it were content. Reader does have a country proxy setting and accepts your own proxy URL, which helps with geo blocks, not with systems that score the browser and IP.
WebScrap is built for that case. Each request is fetched from our own servers and rendered in a browser when you ask for it. A request that still fails is not billed.
A worked example: 100,000 pages a month
Take a price intelligence pipeline that reads 100,000 pages a month. We assume 20 percent of them, 20,000 pages, sit behind anti bot protection and the rest are open. The share is an assumption for the example, so replace it with your own failure log.
| Setup | Open pages (80,000) | Protected pages (20,000) | Result |
|---|---|---|---|
| Jina Reader only | Markdown, billed in tokens | Challenge text or errors | About 20,000 missing or bad records a month |
| WebScrap only | 50,000 requests included, 30,000 on the next plan | Fetched from our own servers | Growth, 149 USD for 250,000 requests |
| Reader plus WebScrap retry | Markdown, billed in tokens | 20,000 requests | Starter, 49 USD, plus your Reader tokens |
The third row is the cheapest way to keep complete data when most of your URLs are open. It keeps the protected 20 percent on a flat plan of 49 USD for 50,000 successful requests. If your protected share is above half, a single route through WebScrap is simpler, because every page then costs one request regardless of length.
A retry setup that keeps both
Send every URL to Reader first. Treat a response as a failure when the status is an error, the body is shorter than a threshold you choose, or the text matches a challenge phrase such as a verification prompt. Send those URLs to the scrape endpoint with browser rendering, and store the result with the route that produced it so you can see the failure rate by domain.
After a week the log shows which domains always fail on Reader. Send those straight to WebScrap and skip the first attempt. That saves time, since the Reader call is a full fetch before it fails. If you would rather return named fields than markdown, the AI web scraper takes a short description and returns JSON, which a model can read with far fewer tokens than a page.
For teams that go on to turn supplier documents into emissions figures, carbon accounting software that classifies invoices picks up where the collection step ends. The full comparison of the two services is on the Jina Reader alternative page.
Questions buyers ask
What are the Jina Reader API rate limits? Jina lists Reader at 20 requests a minute without a key, 500 with a free or paid key and 5,000 on a premium key. Search is blocked without a key, then 100 and 1,000 requests a minute. Limits also apply per token minute and per IP.
How does Jina Reader charge for tokens? Reader counts the tokens in the response it returns, so a longer page costs more. Failed requests are not charged. Search uses a fixed cost per request that starts at 10,000 tokens. OCR mode costs 40 times more tokens.
Does Jina Reader work on Cloudflare protected sites? Not reliably. Jina states that Reader does not actively bypass website defenses or anti bot systems, and that a paid key does not grant access to extra sites. On a protected page you may receive a challenge instead of the content.