Jina AI Search Foundation

Gotchas

• 10M one-time token grant

  Every new API key receives exactly 10 million free tokens shared across all Jina APIs (Embeddings, Reranker, Reader, Classifier, DeepSearch); once exhausted, paid top-up is required.

• 100 RPM / 100K TPM (Embeddings & Reranker)

  Free API keys are capped at 100 requests per minute and 100,000 tokens per minute on the Embedding and Reranker APIs.

• 2 concurrent requests on free tier

  Free keys allow only 2 concurrent in-flight requests across all Jina APIs, limiting parallelism in batch workloads.

• 25 RPM / 25K TPM (Classifier API)

  The Classifier train and classify endpoints are limited to 25 RPM and 25,000 TPM on free keys.

• 50 RPM (DeepSearch)

  Free keys are limited to 50 requests per minute on the DeepSearch endpoint.

• jina-embeddings-v4 noncom only

  The jina-embeddings-v4 model is released under the Qwen Research License (CC BY-NC), which prohibits commercial or production use; only v3 and older models can be used commercially via the API.

• jina-embeddings-v4 intentionally throttled

  Because Jina cannot monetize v4, they intentionally cap its API throughput to control infrastructure costs; high-volume use is explicitly unsupported and callers are directed to self-host.

• Cold starts on serverless models

  Low-traffic models are offloaded during idle periods; the first request after inactivity triggers a warm-up delay of several seconds.

• Token pool shared across all APIs

  The 10M free tokens are shared across Reader, Embeddings, Reranker, Classifier, and DeepSearch — Reader searches alone cost 10,000 tokens per request, so heavy multi-product use can drain the pool quickly.