Jina AI
Listed at https://jina.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Embeddings v3: $0.02/M tokens (10K free); Reranker v3: $0.018/M tokens (10K free); Reader: $0.02/M tokens (free up to 1M tokens); CLIP-v1: $0.02/M tokens | per million tokens |
| Output | Flat per-million-token billing, no tier-based markup; batch discount of 50% available on all endpoints | per million tokens |
🤖 Supported Models (9)
✨ Pros
- ✓jina-embeddings-v3: 570M-param SOTA open-source embedding (8K context + Matryoshka dimension reduction: 32/64/128/256/512/768/1024 switchable)
- ✓Reader API: one line of code to convert any URL to clean Markdown — the de-facto RAG data collection primitive
- ✓jina-reranker-v3: multilingual reranker supporting 100+ languages, outperforms Cohere Rerank-3 on Chinese/English mixed retrieval
- ✓CLIP-v1 multimodal embeddings: native text-image cross-modal retrieval, 89 languages
- ✓RAG-native endpoint trio: Embeddings + Reranker + Reader — can be called independently or pipelined
- ✓10K tokens/month free tier, permanent, no credit card required
- ✓Multiple open-source base models (Apache 2.0): jina-embeddings-v2-base-* 4 variants, commercial use OK
- ✓OpenAI-compatible API endpoint, near-zero migration cost
⚠️ Cons
- ×Embedding/reranker/reader only — no LLM/chat capability
- ×China access requires proxy (AWS US-East + EU-Frankfurt, no mainland edge nodes)
- ×Embeddings v3 max 8192 tokens per request; longer documents need chunking
- ×Reader API has limited support for SPA/JavaScript-rendered pages; complex web apps still need headless browser fallback
- ×No Chinese documentation, English prompts give best results
- ×No formal affiliate program (API link is just UTM source tracking)
🎯 Best For
RAG pipeline embedding + reranking + data collection; multilingual (zh/en/de/fr/es/ja/ko) retrieval; long-document chunking with dimension reduction; native LlamaIndex/LangChain/AutoGen integration
💰 Pricing & Plans
| Plan | Tokens | Price | Effective Rate |
|---|---|---|---|
| Toy Experiment | 10M free | $0 | Free (CC-BY-NC, non-commercial) |
| Prototype Development | 1B | $50 | $0.05 / 1M tokens |
| Production Deployment | 11B | $500 | $0.045 / 1M tokens |
🔧 API & Developer Experience
- •Endpoints: Separate RAG-focused REST endpoints — /v1/embeddings, /v1/rerank, /v1/reader, /v1/search — designed around search & retrieval pipelines rather than chat.
- •OpenAI-compatible: Embeddings endpoint mirrors the OpenAI /v1/embeddings shape, so switching from text-embedding-3/ada is a near drop-in with just a base-URL + key change.
- •Reader & Search: Prefix any URL with r.jina.ai to return clean LLM-ready Markdown; s.jina.ai returns a SERP for web search — both usable with a single GET request.
- •MCP Server: mcp.jina.ai exposes Reader + Search as MCP tools, letting Claude/Cursor/GPT agents fetch live web content inside a single session.
- •Batch Discount: 50% batch discount on all endpoints — submit asynchronous batch jobs for bulk embedding, reranking, or document ingestion at half the per-token cost.
- •SDK & Tools: SDKs for Python and Node plus native integrations with LlamaIndex, LangChain, and Haystack; vector DB adapters (Qdrant, Weaviate, Chroma) built in.
- •Multilingual & Dynamic Dims: Matryoshka output lets you request 32→2048 dimensions from one model; 100+ languages supported without language-specific models.
🔍 RAG-Native Search Foundation
Jina AI is a search-foundation company: it sells the three primitives a retrieval system needs — embeddings, rerankers, and a URL reader — rather than a general chat model. The current line-up is the v5 family: jina-embeddings-v5-omni-small (1.7B, 32K context, text/image/audio/video) shares one embedding space across every modality ('one embedding for all'), so a text query can retrieve video, image, or audio documents without separate models; v5-text-nano (239M) targets edge deployment at 8K context. jina-reranker-v3.5 (2026-07) brings listwise reranking over 131K-token inputs in 100+ languages. Because every endpoint draws from one token pool, a full RAG pipeline — embed, retrieve, rerank, read — is priced and billed uniformly. Open-weight (Apache/AI models) versions let teams self-host for production.
🌐 China Access & Latency
Jina AI's API is hosted on AWS US-East and EU-Frankfurt with no mainland-China edge nodes, so direct access from mainland China requires a stable proxy or VPC/secondary routing. Latency from China typically lands in the 300ms–1.5s range depending on proxy quality — acceptable for batch ingestion but noticeable for real-time rerank inside a search loop. For teams serving China-based users, the practical options are (a) route through a self-hosted aggregator or edge, or (b) self-host Jina's open-source embedding/reranker weights on a domestic GPU. The API itself is pay-as-you-go by token, so there is no regional pricing penalty — but curl-level calls from Chinese IPs will fail without a working proxy.