How does NumStack calculate AI token costs?

NumStack uses real published pricing from OpenAI, Anthropic, Google, and other AI providers. Enter your typical usage (input tokens, output tokens, requests per day) to see projected monthly costs across all models.

Why do different AI models have different token prices?

Different models have different computational costs. Smaller models like GPT-4o mini or Haiku are cheaper but less capable. Larger models like GPT-4o or Claude Sonnet are more powerful but cost more per token.

How can I reduce my AI API costs?

Use smaller models for simple tasks, cache repeated prompts, trim output length (output tokens cost 3-5x more than input), and batch requests instead of real-time when latency allows.

← Back to NumStack

Vector Database Cost Calculator

Compare Pinecone, Weaviate, Qdrant, Supabase pgvector, and Chroma. Know your vector DB costs before you build.

How vector database pricing works

Vector databases charge for two things: storage (keeping your embeddings) and queries (similarity searches at runtime). Storage is cheap at small scale but grows linearly. Queries dominate costs at high traffic.

You also pay your embedding provider once to convert documents into vectors. If you ever switch models, you pay that fee again — so picking the right model upfront matters.

Pinecone vs Weaviate vs Qdrant vs Supabase vs Chroma

Pinecone — Fully managed, zero ops. Best for teams that want to ship fast. Serverless billing fits bursty workloads.
Weaviate — Excels at hybrid search (BM25 + vector). Great for semantic search where keyword context matters.
Qdrant — Best price/performance at scale. Supports scalar and product quantization to shrink storage 4–32×.
Supabase pgvector — Best for Postgres users. Vectors live alongside your relational data. No separate infra.
Chroma — Open-source, free to use. You own the infra. Great for prototyping or private deployments.

How to reduce vector database costs

•Reduce dimensions — Use text-embedding-3-small with 512 or 768 dimensions instead of 1536. Quality drops slightly but costs fall 3×.
•Quantization — Qdrant supports scalar (int8) and product quantization, reducing storage 4–32× with minimal accuracy loss.
•Chunk wisely — Smaller chunks mean more vectors. Aim for 200–500 tokens per chunk with overlap rather than sentence-level splitting.
•Cache frequent queries — If the same queries repeat, cache results in Redis to skip vector DB calls entirely.
•Use free tiers — Pinecone has a free tier (100K vectors). Supabase free tier includes 500MB. Good for prototyping.

🔌 Building a RAG pipeline? Also check the Agent Workflow Cost Calculator to estimate your LLM costs alongside vector DB spend.

Vector database vendors

Some links below may be affiliate links.

Pinecone

A fully managed vector database for building search and retrieval applications.

Visit Pinecone →

Weaviate

An open-source vector database with built-in hybrid search, available self-hosted or managed.

Visit Weaviate →

Qdrant

An open-source vector database and search engine, available self-hosted or as a managed cloud service.

Visit Qdrant →

Zilliz Cloud

A managed cloud service for Milvus, an open-source vector database.

Visit Zilliz Cloud →

How this calculator works

Storage is calculated from your document count and embedding dimensions: each vector takes 4 bytes per dimension (float32), and the calculator adds roughly 30% overhead for metadata and indexing to get total gigabytes. Embedding cost is a one-time charge for turning your documents into vectors, based on your chosen embedding model's current per-token price shown in the calculator.

For each provider (Pinecone, Weaviate, Qdrant, Supabase pgvector, and self-hosted Chroma), monthly cost is that storage figure times the provider's current per-GB rate, plus your monthly query volume times its current per-query or per-operation rate — using each provider's published pricing model, since they charge differently (serverless-per-GB, included-storage-plus-overage, or flat base fee plus usage). The calculator recommends a provider by combining lowest total cost with use-case fit — for example, favoring hybrid search support for semantic search, or higher-throughput options for recommendation workloads.

Storage size

storageGB = (documentCount × dimensions × 4 bytes × 1.3) ÷ 1e9

monthlyCost = storageGB × provider's per-GB rate + queryVolume × provider's per-query rate

How to use it

1.Enter how many documents (or chunks) you plan to store as vectors.
2.Pick your embedding model — this sets both the vector dimensions and the one-time embedding cost.
3.Enter your expected monthly query volume.
4.Select your primary use case (RAG, semantic search, recommendations, or deduplication) so the recommendation reflects what matters most for it.
5.Compare the provider table for monthly cost, then read the recommendation for which provider best fits your numbers and use case.

Common mistakes

⚠Using default 1536 dimensions without thinking — larger embedding models produce more dimensions and therefore more storage — a smaller-dimension model can cut storage cost 2-3x with only a modest quality tradeoff for many use cases.
⚠Forgetting re-embedding cost — if your source documents update regularly, you'll pay the embedding cost again for changed content — a one-time embedding estimate understates cost for frequently updated corpora.
⚠Ignoring query volume until it's large — storage cost dominates at low query volume, but query cost can overtake it fast for high-traffic search or recommendation features — check both line items, not just storage.
⚠Not matching the provider to the use case — the cheapest provider by raw cost isn't always the best fit — hybrid search, existing Postgres infrastructure, or self-hosting tolerance should factor into the choice alongside price.

Frequently asked questions

How much does a vector database typically cost?▾

It depends entirely on your document count, embedding dimensions, and query volume — the calculator prices all five providers against your specific numbers using their current published rates, rather than a single fixed answer, since the right number for a small prototype and a production RAG pipeline can differ by orders of magnitude.

Which vector database is cheapest?▾

It varies by workload size and shape — self-hosted options are typically cheapest if you already have infrastructure and ops capacity, while serverless managed options are often more cost-efficient at low volume since you only pay for what you use. Compare the provider table above for your specific inputs.

How much does it cost to embed my documents?▾

Embedding cost is your total document count times the selected embedding model's current per-token price, shown in the calculator. Open-source embedding models run locally and add no per-token API cost, only compute you already own or rent.

What is the practical difference between the providers compared here?▾

Pinecone is fully managed with minimal ops overhead. Weaviate includes native hybrid BM25-plus-vector search, useful for semantic search quality. Qdrant offers strong price/performance with quantization support at scale. Supabase pgvector fits well if you already run Postgres. Chroma is open-source and free to self-host, trading a lower bill for your own ops work.

How is storage size calculated?▾

Each vector uses 4 bytes per dimension (float32 precision), multiplied by your document count, plus roughly 30% overhead for metadata and indexing. Reducing dimensions or enabling quantization (where a provider supports it) reduces this figure directly.

Does the calculator account for quantization?▾

Quantization support is reflected qualitatively in provider notes (Qdrant, for example, supports it) but the cost math uses full-precision storage — if you plan to enable quantization, treat the storage cost shown as an upper bound.