Editor’s note: This Jina AI review is based on my hands-on test of the Jina Search Foundation API (Reader, Embeddings with jina-embeddings-v5, Reranker v3.5 and DeepSearch) on a new API key, in November 2026. I weighed token pricing and rate limits, the embedding models, reranking, Reader and Search for web grounding, DeepSearch, licensing and the Elastic acquisition, and how it compares with OpenAI, Cohere, Voyage AI, Google Gemini and Firecrawl.
Quick verdict: I recommend Jina AI for developers building RAG and agent pipelines who want multilingual embeddings, a reranker and a URL-to-Markdown reader on one key, billed from one token balance at $0.05 per 1M tokens. The trade-off is licensing: the free 10M-token grant and every open-weight v3, v5 and current reranker model are non-commercial (CC BY-NC 4.0), so commercial self-hosting means a Jina On-Prem license from Elastic Sales, with no public price.
Jina AI has been owned by Elastic since October 9, 2025, but the self-serve API this Jina AI review covers still runs Reader, Search, Embeddings, Reranker, Classifier and Segmenter from one key and one balance. Its jina-embeddings-v5-text-small packs a 32,768-token context and 119 languages into 677M parameters.
In this review, I’ll take a closer look at Jina AI’s pricing, models, and licensing, so you can see exactly whether it fits your retrieval stack or whether OpenAI, Cohere or Voyage AI is the better buy.
Key Takeaways
- One key, one balance: six APIs spend the same tokens, at $0.05 per 1M on the $50 pack (1B tokens) and $0.045 on the $500 pack (11B tokens).
- Small, long-context models: jina-embeddings-v5-text-small fits 1,024 dimensions, a 32,768-token context and 119 languages into 677M parameters; the omni models add images, audio, video and PDFs.
- Web grounding bills differently: Reader runs keyless at 20 requests a minute, but a keyed Reader call bills its output tokens and every Search request costs from 10,000 tokens.
- Free is not commercial: the open weights and the 10M-token grant are CC BY-NC; commercial self-hosting needs a Jina On-Prem quote from Elastic Sales.
- Throughput and compliance gaps: a paid key caps Embeddings and Reranker at 500 RPM and 2M TPM below the $500 pack, and jina.ai publishes no SOC 2 or ISO 27001 report.
Pros and Cons
Pros
- One key and one token balance for six retrieval APIs.
- $0.05 per 1M tokens on the $50 pack; failed requests are not charged.
- v5-text-small: 677M parameters, 32K context, 119 languages.
- Reader works without a key at 20 requests a minute.
- OpenAI-compatible Embeddings and DeepSearch endpoints.
Cons
- CC BY-NC weights: commercial self-hosting needs a quoted license.
- 500 RPM and 2M TPM until you buy the $500 pack.
- Reader bills output tokens; each Search starts at 10,000 tokens.
- No SOC 2 or ISO 27001 report published on jina.ai.
Jina AI Pricing: How Much Does It Cost?
Every new Jina API key starts with 10M free tokens; after that, prepaid packs fill one balance shared by every Search Foundation API, at $0.05 per 1M tokens on the $50 pack.
- No key ($0): for Reader only, at 20 requests a minute per IP; Search, Embeddings and Reranker need a key.
- Free key ($0, 10M tokens): for prototypes, labeled “Toy Experiment” for non-commercial use only (CC BY-NC); Reader runs at 500 RPM and Search at 100 RPM.
- Standard pack ($50 for 1B tokens): for a first production workload, at $0.050 per 1M tokens with priority technical support. Embeddings and Reranker rise to 500 RPM and 2M TPM.
- Premium pack ($500 for 11B tokens): for higher-throughput production, at $0.045 per 1M tokens with advanced key management, support within 24 hours and a one-hour integration consultation. Reader rises to 5,000 RPM and Search to 1,000 RPM.
- 110B pack ($5,000 for 110B tokens): for volume buyers, at the same $0.045 rate and premium limits.
- Elastic Inference Service and Jina On-Prem (custom): for Elastic Cloud customers or commercial self-hosting; On-Prem is an annual license with no public price.
| Option | Price | Tokens | Per 1M tokens | Embeddings/Reranker limit |
|---|---|---|---|---|
| No key | $0 | None (Reader only) | n/a | Blocked |
| Free key | $0 | 10M (non-commercial) | $0 | 100 RPM, 100K TPM |
| Standard pack | $50 | 1B | $0.050 | 500 RPM, 2M TPM |
| Premium pack | $500 | 11B | $0.045 | 5,000 RPM, 50M TPM |
| 110B pack | $5,000 | 110B | $0.045 | 5,000 RPM, 50M TPM |
| Elastic Inference Service / On-Prem | Custom | Contract | Custom | Contract |
Stripe takes cards, Google Pay and PayPal, and an optional auto top-up refills the key at a low-balance threshold. Failed requests are not charged, and both balance and rate limits belong to the key, not the account.
Token counting by API:
- Embeddings and Reranker: input tokens.
- Reader: the output tokens it returns.
- Search: a fixed charge per request, starting at 10,000 tokens.
- Reader with jina-ocr-v1: 40x the tokens.
At the 10,000-token minimum, the free grant covers at most 1,000 Search requests, and $50 covers up to 100,000.
Is Jina AI Good Value for Money?
- Embeddings: Jina’s $0.05 per 1M sits above OpenAI text-embedding-3-small and Voyage AI voyage-4-lite ($0.02 each) but below voyage-4 ($0.06), Cohere Embed 5 Fast ($0.08) and Gemini Embedding ($0.15).
- Reranking: $0.05 per 1M matches Cohere Rerank 5 Pro and Voyage AI rerank-2.5; their Fast and lite tiers charge $0.02.
- Web reading: Firecrawl‘s Hobby plan is $19 a month for 5,000 credits at 1 credit a page; Jina’s Reader runs keyless at 20 RPM and bills tokens with a key.
- Model choice: the rate is flat, so v5-omni and the 0.6B reranker cost the same per token as v5-text-nano.
Jina is good value when one balance replaces three vendors: an embedding API, a reranker and a scraper. For a text-only corpus at volume, OpenAI text-embedding-3-small costs less than half as much per token.
Reviewer’s Notes: I recommend prototyping on the free 10M tokens and reading each Embeddings response’s usage object before buying the $50 pack. Move to the $500 pack only when Embeddings or Reranker needs more than 500 RPM or 2M TPM; self-hosted weights need Elastic Sales, not a bigger pack.
My Experience With Jina AI
I ran the whole test on one auto-generated key, which covered Reader, Search, Embeddings, Reranker and DeepSearch.
Getting My API Key and Free Tokens
I opened the Embeddings page on jina.ai, which generated an API key automatically and loaded it with 10M free tokens for non-commercial use. I checked the balance in the API Key & Billing tab. Jina’s API keys do not expire, and a paid pack tops up the same key.
Voyage AI gives 200M free tokens per account, far more than Jina’s 10M per key, but it has no URL reader.
Turning Web Pages Into Markdown With Reader
I prefixed a public page URL with r.jina.ai and sent it with no key, which Jina allows at 20 requests a minute per IP. Jina Reader strips navigation, footers and ads, returns the main content as Markdown and reads PDFs natively. Then I added my key, which raises the Reader limit to 500 RPM and bills the output tokens, and set the request headers:
- X-Engine: I switched between direct (a plain HTTP request, the fastest engine) and the default headless browser, which runs client-side JavaScript.
- X-Target-Selector and X-Remove-Selector: I kept one section of a page and dropped another.
- X-Token-Budget and X-Timeout: I capped the tokens and the wait for each request.
- X-With-Generated-Alt: I turned on generated captions for images.
- Accept: application/json: I asked for URL, title and content as separate JSON fields.
A repeat request for the same URL within 5 minutes comes from cache unless X-No-Cache is set, and Reader cannot read local HTML files. I also sent queries to s.jina.ai, which returns the top five results with page content. For whole-site jobs, Firecrawl adds endpoints that scrape and map an entire site.
A paid key raises throughput, but Reader does not get past sites that block it, on any tier.
Embedding and Reranking a Small Corpus
I sent POST requests to the embeddings endpoint on api.jina.ai with model jina-embeddings-v5-text-small and task retrieval.passage for documents or retrieval.query for questions. I used dimensions to truncate the 1,024-dimension vectors (Matryoshka) and turned on late_chunking so each chunk vector keeps whole-document context. Jina says this schema matches OpenAI’s text-embedding-3-large, so moving an OpenAI client usually takes a new base URL, key and model name.
Every Embeddings response includes a usage object with the exact tokens charged. I then passed the top candidates to jina-reranker-v3.5 on the same key.
Vectors from different model families are not comparable, so switching families means re-embedding the corpus. Jina stores no vectors; they still need a database such as Pinecone, Milvus or Chroma.
Topping Up and Managing Keys
I opened the top-up widget, which offers the $50 and $500 packs with x10 variants, plus an auto top-up threshold. Any paid key can move its remaining tokens to a new key during rotation, through Transfer Tokens or Revoke with transfer; a free key has no paid balance to move.
Reviewer’s Notes: Run your real documents and pages through the free key and watch the API Key & Billing tab before buying a pack. A Search request starts at 10,000 tokens and Reader bills every output token, so short embedding queries are a poor guide to what web grounding will cost.
Jina Embeddings v5: Small Multilingual Models With Long Context
Released in February 2026, jina-embeddings-v5-text-small has 677M parameters, 1,024 dimensions and a 32,768-token context, covers 119 languages and carries four LoRA task adapters: retrieval, text-matching, classification and clustering. Jina reports 67.0 on MMTEB and 71.7 on MTEB English, and says it matches the 3.8B jina-embeddings-v4 on its retrieval average (63.28 vs 63.62) at 5.6x smaller. The 239M v5-text-nano scores 65.5 on MMTEB by Jina’s figures.
| Model | Parameters | Dimensions | Context | Inputs | Weight license |
|---|---|---|---|---|---|
| jina-embeddings-v5-text-small | 677M | 1,024 (Matryoshka 32-1,024) | 32,768 | Text | CC BY-NC 4.0 |
| jina-embeddings-v5-text-nano | 239M | 768 (Matryoshka 32-768) | 8,192 | Text | CC BY-NC 4.0 |
| jina-embeddings-v5-omni-small | About 1.74B | 1,024 | 32,768 | Text, image, audio, video, PDF | CC BY-NC 4.0 |
| jina-embeddings-v5-omni-nano | About 1.04B | 768 | 8,192 | Text, image, audio, video, PDF | CC BY-NC 4.0 |
| jina-embeddings-v4 | 3.8B | 2,048 dense or 128 multi-vector | 32,768 | Text, image | Qwen Research (non-commercial) |
| jina-reranker-v3.5 | 0.6B | n/a (reranker) | 131,072 | Text | CC BY-NC 4.0 |
The omni models produce text vectors identical to v5-text-small and -nano, so images, audio, video and PDFs can join an existing text index without re-embedding the text. Video is sampled as 32 uniform frames, and a 600×600 image costs about 363 tokens on v5-omni-small against 4,840 on v4 and 16,000 on jina-clip-v2. Requests accept images up to 5 MB and one PDF of up to 8 MB, while OpenAI’s text-embedding-3 models are text only.
Jina assigns each model a job:
- jina-embeddings-v5-text-small: text retrieval, including multilingual and long documents.
- jina-embeddings-v5-text-nano: when latency, cost or edge hardware matter.
- jina-embeddings-v5-omni-small or -omni-nano: images, audio, video or PDFs.
- jina-code-embeddings (0.5B and 1.5B): source code.
For most pipelines, Jina recommends a dense v5 model plus a reranker over multi-vector retrieval.
Skip jina-embeddings-v4 for anything you ship: it is free on the API only because its Qwen Research license bars commercial use, it is throttled, and Jina calls it not suitable for production. It is not offered on Elastic Inference Service or Jina On-Prem.
Best for: multilingual and long-document retrieval with a path to mixed media. Skip if: you need open weights you can ship commercially without a license deal.
Reranker, Search and ReaderLM: Retrieval Beyond Embeddings
jina-reranker-v3.5, released in July 2026, is a 0.6B listwise reranker with a 131,072-token context, and Jina says it runs 1.22x to 1.56x faster than jina-reranker-v3. Jina’s other models:
- jina-reranker-v3: the 0.6B predecessor, with a 131K context and 100+ languages.
- jina-reranker-m0: a 2.4B multimodal reranker for visual documents, with a 10K context.
- jina-colbert-v2: a 559M late-interaction model with an 8,192-token context and 89 languages.
- ReaderLM-v2: a 1.54B model that turns raw HTML into Markdown or JSON, up to 512K tokens of combined input and output, in 29 languages.
- jina-ocr-v1: Reader’s OCR option for scanned pages, at 40x the token cost.
- jina-vlm: a 2.4B vision-language model for scanned or image-heavy documents.
Reader hands x-json-schema or x-instruction to ReaderLM-v2, which returns named fields as JSON instead of a whole Markdown page.
Jina’s own latency figures show reranking time climbing with document length: about 150 ms for 100 documents of 256 tokens with a 64-token query, about 3.5 seconds with 4,096-token documents, and about 7 seconds with a 512-token query on 4,096-token documents. Jina lists average latency of 7.9 seconds for Reader and 2.5 seconds for Search.
On top of the per-key limits, Jina caps each IP at 10,000 requests per 60 seconds, and over-limit requests return HTTP 429. Classifier and Segmenter share the same balance but have no published rate limits.
Cohere Rerank 5 Fast and Voyage AI rerank-2.5-lite cost $0.02 per 1M tokens against Jina’s $0.05, so the Jina reranker is not the cheapest standalone pick. It earns its place when Reader and Search run on the same key.
DeepSearch: An OpenAI-Compatible Research Agent API
DeepSearch takes OpenAI chat-completions requests at deepsearch.jina.ai with model jina-deepsearch-v1, then searches, reads and reasons in a loop until it reaches an answer or spends its budget. Its controls:
- reasoning_effort: low, medium or high, a preset of budget_tokens and max_attempts that explicit values override.
- team_size: parallel agents sharing one token budget.
- Domain lists: good, bad and only lists that steer which sites it searches.
- Structured output: answers shaped by a JSON schema.
- Attachments: webp, png or jpeg images and txt or PDF files up to 10 MB, as data URIs.
- stream: strongly recommended, since non-streaming calls can hit 524 timeouts.
Other settings cover search and answer language, a no_direct_answer flag and the maximum number of returned URLs. DeepSearch bills from the same token balance with no per-request price, so budget_tokens is the direct cost cap. Set it explicitly when raising team_size, because every parallel agent draws on that one budget.
My verdict: a useful research primitive inside your own agent, not a finished research product for end users.
Licensing, Elastic and Deployment: What Changed After the Acquisition
Elastic completed its acquisition of Jina AI on October 9, 2025, and jina.ai now offers three paths:
- Search Foundation API: self-serve, pay-as-you-go tokens on a Jina key.
- Elastic Inference Service: Jina models inside Elastic Cloud, billed through the Elastic subscription.
- Jina On-Prem: a commercial self-hosting license Elastic has sold since August 10, 2026, priced annually rather than per token, with no public price.
The v3 and v5 embedding weights, the current rerankers, jina-colbert-v2 and ReaderLM-v2 are CC BY-NC 4.0 on Hugging Face, so the free weights cannot run in a commercial product, and token purchases cover only the hosted API. Jina On-Prem ships as fully offline Docker containers with no license server and exposes Elastic Inference Service, OpenAI, Cohere, Voyage AI and Gemini API schemas.
Jina marks 29 models as air-gapped deployable and lists models on AWS Marketplace (SageMaker), Azure Marketplace and Google Cloud Marketplace for deployment in your own cloud account. Elastic lists jina-embeddings-v5-text-small, v5-omni-small, v5-omni-nano and v3 on Elastic Inference Service, usable with the semantic_text field type, so teams already on Elasticsearch can skip a separate Jina key. Purchase orders, negotiated contracts, fine-tuning on proprietary data and dedicated capacity all go through Elastic Sales.
Jina states it never trains its models on API requests, inputs or outputs, stores them only as needed to run the service, and encrypts data in transit and at rest, with processing under Elastic’s Customer DPA. jina.ai cites OWASP Top 10 alignment, external penetration testing and customer isolation, but publishes no SOC 2 or ISO 27001 report.
How Does Jina AI Compare to Competitors?
Jina AI puts embeddings, reranking, a URL reader and web search on one balance, but OpenAI text-embedding-3-small and Voyage AI voyage-4-lite both charge $0.02 per 1M tokens against Jina’s $0.05 for text embeddings.
| Vendor | Text embeddings per 1M tokens | Reranking per 1M tokens | Main buying distinction |
|---|---|---|---|
| Jina AI | $0.050 standard; $0.045 premium | $0.050 standard; $0.045 premium | Reader and Search share the balance |
| OpenAI | $0.02 small; $0.13 large | None | Text only, no reranker |
| Cohere | $0.08 Fast; $0.12 Pro | $0.02 Fast; $0.05 Pro | Multimodal, 128K context, private deployment |
| Voyage AI | $0.02 lite; $0.06 voyage-4; $0.12 large | $0.02 lite; $0.05 rerank-2.5 | 200M free tokens and domain models |
| Google Gemini | $0.15 Embedding; $0.20 Embedding 2 text | None | Embedding 2 covers text, image, audio and video |
- OpenAI: text-embedding-3-small costs $0.02 per 1M tokens and text-embedding-3-large $0.13, both text only, with no reranker. It fits text workloads that will never touch Reader or Search.
- Cohere: Embed 5 Pro costs $0.12 per 1M text tokens ($0.40 per 1M image tokens) and Embed 5 Fast $0.08, with 128K context, multimodal input and 100+ languages. It fits enterprises that want private deployment on Amazon SageMaker or Microsoft Foundry and accept proprietary weights.
- Voyage AI: 200M free tokens per account and domain models for code, finance and law, with voyage-4 at $0.06 per 1M tokens. It fits teams that want the biggest free allowance or domain-tuned retrieval, though it has no URL reader.
- Google Gemini: Gemini Embedding costs $0.15 per 1M text tokens, and Gemini Embedding 2 charges $0.45 for images, $6.50 for audio and $12.00 for video per 1M tokens. It fits teams that want one multimodal model and accept a higher rate.
- Firecrawl: 1,000 free credits a month, Hobby at $19 for 5,000 credits and Standard at $99 for 100,000, at 1 credit per scraped page, with endpoints that scrape and map an entire site. A scrape of a blocked or missing page still costs a credit, while Jina does not charge failed requests. Apify fits when you need ready-made scrapers for specific sites.
Jina wins when one key must cover embeddings, reranking and web grounding for multilingual or long documents. Among the Jina AI alternatives, OpenAI wins on price for plain text, Voyage AI on free allowance, and Cohere when private deployment on SageMaker or Microsoft Foundry matters more than per-token price.
How I Reviewed Jina AI
I created a new API key on the Embeddings page, converted pages with Reader with and without a key, ran Search queries, embedded a small corpus with jina-embeddings-v5-text-small, reranked it with jina-reranker-v3.5 and sent a DeepSearch request. I worked through the token packs, rate limits and token-counting rules, checked the model licenses and Elastic paths, and compared Jina with OpenAI, Cohere, Voyage AI, Google Gemini and Firecrawl. I gave the most weight to:
- Pricing and value: packs and token counting.
- Models: context, dimensions, media.
- Reranking: models and latency.
- Web grounding: Reader and Search.
- Developer experience: schemas and keys.
- Licensing and deployment: hosted versus self-hosted.
- Data handling: training policy and DPA.
- Alternatives: cheaper embeddings and whole-site scraping.
Prices current as of November 2026.
Jina AI Review: Should You Use Jina AI for Embeddings and RAG?
Yes, use Jina AI if you are building RAG, search or agent pipelines that need multilingual or long-context embeddings, a reranker and URL-to-Markdown reading on one key, and a hosted API at $0.05 per 1M tokens (or Jina models inside Elastic Cloud) works for you. Skip it if:
- You embed plain text at volume: OpenAI text-embedding-3-small costs $0.02 per 1M tokens.
- You want the largest free allowance: Voyage AI gives 200M tokens per account.
- You need whole-site scraping: Firecrawl has endpoints that scrape and map an entire site.
- You need commercial weights to self-host: Jina’s free weights are CC BY-NC, so self-hosting outside the cloud marketplaces means a quoted Jina On-Prem license.
- Procurement needs a published SOC 2 or ISO 27001 report: jina.ai publishes neither.
Before buying a pack, measure your tokens per Reader page and per Search request: at the 10,000-token minimum, $50 covers at most 100,000 searches. If Embeddings or Reranker traffic peaks above 500 RPM or 2M TPM, budget for the $500 premium pack, the entry point to premium limits of 5,000 RPM and 50M TPM.
Start a non-commercial prototype on the free 10M tokens at Jina AI.
FAQ
Is Jina AI free?
Partly. Every new key gets 10M free tokens, labeled “Toy Experiment” for non-commercial use only, and Reader works with no key at 20 requests a minute per IP. After the grant, paid packs cost $50 for 1B tokens or $500 for 11B tokens, and commercial traffic belongs on those packs.
Is the Jina embeddings API compatible with OpenAI?
Yes. Jina says its embeddings endpoint matches the request and response schema of OpenAI’s text-embedding-3-large, so migration usually means a new base URL, key and model name. DeepSearch also accepts OpenAI chat-completions requests with model jina-deepsearch-v1. Re-embed your corpus when you switch model families, because schema compatibility does not make vectors from different models comparable.
Can I use Jina models commercially?
Yes, through the paid API, Elastic Inference Service, the AWS, Azure and Google Cloud marketplaces, or a Jina On-Prem license. The free Hugging Face weights for v3, v5, the current rerankers and ReaderLM-v2 are CC BY-NC 4.0, non-commercial only. jina-embeddings-v4 stays non-commercial under its Qwen Research license.
Does Jina AI train on my data?
No. Jina states that API requests, inputs and outputs are never used to train its models. It stores inputs and outputs only as needed to provide the service, encrypts data in transit and at rest, and processes it under Elastic’s Customer DPA. jina.ai cites OWASP Top 10 alignment and external penetration testing but publishes no SOC 2 or ISO 27001 report.
Is Jina AI owned by Elastic?
Yes. Elastic completed its acquisition of Jina AI on October 9, 2025. The self-serve pay-as-you-go API still runs at jina.ai in November 2026, Elastic Inference Service runs Jina models inside Elastic Cloud, and Elastic Sales handles Jina On-Prem licenses and enterprise contracts.
