Jina Search and DeepSearch: Endpoints, Pricing, Owner

Jina search is a set of web APIs from Jina AI: s.jina.ai searches the web, r.jina.ai reads a page, and DeepSearch repeats search, read and reasoning steps until it can answer a hard question with sources. Jina AI is now owned by Elastic, and the APIs are still live (Jina AI, About us, 2026; Jina DeepSearch page, 2026).

Key Takeaways

• Elastic acquired Jina AI on 9 October 2025. The Jina APIs still run, with a self-serve token price list (Jina AI, About us, 2026).
• DeepSearch is one OpenAI-compatible endpoint, deepsearch.jina.ai/v1/chat/completions, with the model name jina-deepsearch-v1 (Jina DeepSearch page, 2026).
• A new API key includes 10 million free tokens for non-commercial use. Paid tokens cost $0.05 per million, or $0.045 per million on the larger pack (Jina DeepSearch page, 2026).
• Jina's own page puts a typical DeepSearch answer at about 500,000 tokens and about 50 seconds (Jina DeepSearch page, 2026).

Who owns Jina AI now, and is the API still live?

Elastic owns Jina AI. Elastic acquired the company on 9 October 2025, and the footer of jina.ai now reads "Elastic © 2026" (Jina AI, About us, 2026). The Jina APIs are still live: the DeepSearch page was serving the endpoint and price list when checked on 30 October 2026.

Jina models are now sold through three paths. The hosted Jina APIs are self-serve and pay as you go with a Jina API key. The Elastic Inference Service runs Jina models inside Elastic Cloud and bills through your Elastic subscription. Jina On-Prem is a commercial licence that Elastic has sold since 10 August 2026 for self-managed or air-gapped setups (Elastic, Jina search models, 2026). This post covers the first path.

Which Jina search APIs exist today?

Jina offers four APIs that matter for web research: Reader, Search, DeepSearch, and the Embeddings and Reranker models. Reader and Search fetch content, DeepSearch writes cited answers, and the models rank and compare text. All of them share one API key and one token balance (Jina DeepSearch page FAQ, 2026).

API	Endpoint	What it does	Rate limit with a free key
Reader	r.jina.ai	Turns a URL into LLM-ready text	500 requests per minute
Search	s.jina.ai	Searches the web and returns LLM-ready text	100 requests per minute
DeepSearch	deepsearch.jina.ai	Searches, reads and reasons in a loop	Not listed in Jina's rate limit table
Embeddings and Reranker	api.jina.ai/v1/embeddings, /v1/rerank	Vectors and result ranking	100 requests per minute

Source: Jina DeepSearch page, rate limit table, 2026. Without a key, r.jina.ai allows 20 requests per minute and s.jina.ai is blocked.

How does the Jina s.jina.ai search endpoint work?

s.jina.ai takes a query in the URL and returns search results as text a language model can read. Add ?site=jina.ai to limit results to one site, and repeat the site parameter for several sites (Jina Reader on GitHub, 2026). Each request costs a fixed number of tokens, starting at 10,000 (Jina DeepSearch page, 2026).

How does the Jina Reader API read a page?

Put https://r.jina.ai/ in front of any URL and Jina returns the page as clean text. Request headers change the output: X-With-Links-Summary adds a list of links, X-With-Images-Summary adds an image list, and X-Target-Selector keeps one part of the page (Jina Reader page, 2026). Reader charges tokens for the text it returns.

What is Jina DeepSearch and how does it work?

Jina DeepSearch is an agent that searches the web, reads pages, and reasons over what it found until its own check accepts the answer. It stops when the answer passes the check or when the token budget or attempt limit runs out (Jina DeepSearch page, 2026).

The cost and speed differ a lot from a normal search call. Jina's page lists a plain language model at about 1,000 tokens and 1 second, a RAG answer (retrieval-augmented generation, an answer built from one round of search results) at about 10,000 tokens and 3 seconds, and DeepSearch at about 500,000 tokens and 50 seconds.

Which DeepSearch settings control quality and cost?

The DeepSearch settings that control quality and cost are reasoning_effort, budget_tokens, max_attempts and team_size. Higher values give deeper research and use more tokens, so set them to match how hard the question is. The reasoning_effort setting is a preset of two of the others (Jina DeepSearch page, 2026).

Setting	What it controls
reasoning_effort	Preset of budget_tokens and max_attempts: low, medium or high
budget_tokens	Maximum tokens for the whole run
max_attempts	How many times it retries after failing its own answer check
team_size	Number of agents that research sub-problems in parallel

Source controls work the same way. boost_hostnames raises priority for listed domains, bad_hostnames blocks them, and only_hostnames restricts reading to them. search_provider: "arxiv" limits all search to arXiv. Jina recommends streaming, because a request without streaming can fail with a 524 timeout error.

How do you call the Jina DeepSearch API?

You call the Jina DeepSearch API the same way as the OpenAI chat completions API, with a different base URL and model name, so existing OpenAI client code needs only two changes. Send a POST request to https://deepsearch.jina.ai/v1/chat/completions with your key in the Authorization header (Jina DeepSearch page, 2026):

curl "https://deepsearch.jina.ai/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $JINA_API_KEY" \
  -d '{
    "model": "jina-deepsearch-v1",
    "messages": [{ "role": "user", "content": "What changed at Jina AI after the Elastic deal?" }],
    "stream": true,
    "reasoning_effort": "medium"
  }'

The last streamed chunk holds the final answer, the list of URLs it visited, and the token usage. Jina publishes the agent code as open source in node-DeepResearch on GitHub.

How much does Jina search cost?

Jina search costs $0.05 per million tokens on the standard pack and $0.045 per million on the premium pack. The same token price covers Reader, Search, DeepSearch and the models. A new key includes 10 million free tokens, limited to non-commercial use under the CC-BY-NC licence (Jina DeepSearch page, 2026).

Pack	Tokens	Price	Per million tokens
Free key	10 million	$0	Non-commercial use only
Prototype Development	1 billion	$50	$0.050
Production Deployment	11 billion	$500	$0.045

The premium pack also raises rate limits, to 5,000 requests per minute for Reader and 1,000 for Search, and adds support within 24 hours. Jina changed its pricing on 6 May 2025, and older auto-recharge settings keep the older price.

At about 500,000 tokens per DeepSearch answer, one answer costs roughly $0.025 on the standard pack. That is an estimate from Jina's own average, and a high effort run can use more.

How does Jina search compare with SearchAPI and Perplexity AI?

Jina returns search results as text for a language model to read, while SearchAPI returns raw search engine results as structured data. The two are different jobs: one gives you text to read, the other gives you data to filter. Many research workflows use both, and Perplexity AI is a third option when you want a finished answer.

• Jina search and DeepSearch: text output ready for a language model, and an agent loop that writes a cited answer.
• SearchAPI: structured results from engines such as Google and Bing, with no answer written for you (CodeWords Search API page).
• Perplexity AI: an answer engine that pairs a language model with live web search (CodeWords Perplexity AI page).

How do you run web research like DeepSearch in a CodeWords workflow?

You describe the research job to CodeWords and Cody builds the workflow, which can run on a schedule, a trigger or a request (CodeWords, 2026). CodeWords has no Jina integration page today (checked 30 October 2026), so build the loop from integrations that are listed: Firecrawl to read pages, Perplexity AI and Search API to search.

A four-step loop copies the DeepSearch pattern:

1. Split the question. Break one research question into three to five sub-questions.
2. Search each sub-question. Run every sub-question through search, in parallel.
3. Read the best results. Extract full page text, not snippets.
4. Write the cited answer. Ask a language model to combine the findings and cite each source URL.

Start from the AI Web Research Assistant template. For a recurring use case, see how to automate competitor monitoring and the competitor news monitor template. For report output, see deep research markdown.

FAQs: Jina AI DeepSearch

How can I get a free Jina API key?

Open the Jina DeepSearch page and a key is created for you automatically, with 10 million free tokens. The free tokens are for non-commercial use only. Once they run out, you top up the same key with a paid pack. One key works for Reader, Search, DeepSearch, Embeddings and Reranker [Jina DeepSearch page, 2026].

Is Deep Search ai free?

Jina DeepSearch is free only up to the free token allowance: 10 million tokens on a new key, for non-commercial use. After that it costs $0.05 per million tokens on the 1 billion token pack, or $0.045 per million on the 11 billion token pack. A single answer averages about 500,000 tokens [Jina DeepSearch page, 2026].

Is it safe to use deep search?

Jina says it never uses API requests, inputs or outputs to train its models. Jina also publishes a security page, and Elastic now owns the company. As with any hosted API, do not send passwords or private customer data in a prompt unless your own policy allows it [Jina DeepSearch page FAQ, 2026].

Does DeepSearch actually work?

It works best on hard questions that need several searches and sources. The agent checks its own answer and retries within a token budget. Jina publishes no independent accuracy benchmark on its page, so test it on your own questions and read the cited URLs before relying on an answer [Jina DeepSearch page, 2026].