Pinecone Bakes Gemini Embedding 2 Into Managed Search at $0.20

Pinecone now runs Google's Gemini Embedding 2 directly inside its Documents API, charging $0.20 per 1 million tokens and eliminating the separate Google Cloud account, API key and embedding pipeline that text-search teams previously maintained. The company announced the integration on November 12, 2026. The move folds a third-party model dependency into the database path — and, as always with first releases, the fine print narrows the pitch.

How does the integration actually work?

Applications now send plain text to Pinecone. On every write and search query, Pinecone invokes the Gemini model through Vertex AI and retries failed calls on the customer's behalf. Embedding usage lands on the Pinecone invoice rather than a separate Google bill — roughly the cost of two US pennies for every million tokens processed.

Pinecone also writes each document's text and embedding together, so semantic and full-text search operate on the same version of a document. For readers new to the concept: embeddings are vectors that represent meaning, letting one field support both similarity-based retrieval and keyword search without the application generating vectors first.

What are the configuration constraints?

Customers select gemini-embedding-2 in a text field's embed configuration at index creation. From there, the options tighten:

• 3,072 dimensions by default, with any integer from 128 to 3,072 supported
• Cosine as the sole similarity metric
• 8,192 tokens maximum per input
• Up to 96 inputs per embedding request
• 10,000 bytes per search query

Model, dimension and metric are fixed once the index exists. That makes them infrastructure decisions, not parameters teams can tune after deployment — a caveat worth weighing against any promise of flexibility.

What do task types control?

Task types govern how the model embeds text for different workloads. General retrieval defaults to retrieval_document for stored text and retrieval_query for searches. Classification, clustering and sentence-similarity tasks treat stored text and queries alike, while other settings cover question answering, fact verification and code retrieval. A single search can override the field's default query task.

Pinecone inserts Gemini's required task instruction automatically, so users submit plain text instead of building model-specific prompts. Those instructions add roughly six or seven tokens to each value or query — and they count toward the model's input limit. The field's write task is fixed at index creation.

What does it cost, and who can use it?

Charges apply to embedding tokens on both writes and queries, tying consumption directly to indexing and search activity. The Builder plan includes 10 million Gemini Embedding 2 tokens per month, doubled to 20 million through December 31, 2026. The model requires a Builder, Standard or Enterprise plan; Starter users are excluded — the free tier does not get this feature at all.

Where does the first release fall short?

The launch is limited to serverless Documents API indexes. BYOC, pod-based and Records API indexes do not qualify. Console and Python SDK 10.3.0 support is live, with Node.js and Go SDK support "coming soon" — a phrase that carries no date.

Existing vector indexes can convert to Documents API indexes without re-embedding, but adding integrated embedding to a converted index is not yet available. The practical workaround: build new Documents API indexes from scratch. Multimodal support is listed as planned for a later release.

That roadmap points beyond the text-only launch. Google's March 10, 2026 model announcement describes Gemini Embedding 2 as natively multimodal, mapping text, images, video, audio and documents into one embedding space. Pinecone's text service covers more than 100 languages. Processing is global: Google can handle each request in any of its regions, with no guarantee of a particular region — a detail that matters for anyone with data-residency requirements, and one Pinecone does not resolve.

For teams operating search infrastructure, the release shifts embedding from a separately managed provider integration into the database itself. Its immediate fit is new text-search deployments on paid plans; global-only processing and the converted-index limitation define the honest scope of the initial service.