Pinecone Pricing (2026): What 1M to 100M Vectors Cost

Pinecone pricing starts at $0 for Starter and $20 a month for Builder, but a modeled 100M-vector corpus costs $11,679.22 a month on Standard at one million whole-namespace queries using its lowest published read rate. For RAG and agent memory, the deciding variable is how much data each retrieval searches, alongside your write volume.

Prices and plan limits were verified against Pinecone's live pricing page on October 5, 2026. The calculations below use its current on-demand rates and Pinecone's own metering formulas. Amounts are in US dollars before tax.

Pinecone is a managed vector database: it stores numerical representations of your content and retrieves similar records. In retrieval-augmented generation, or RAG, those retrieved records supply context to a language model. An agent memory system does the same with saved information from previous work. The database bill covers that storage and retrieval; generating embeddings and answering with a language model can add separate costs.

Pinecone Pricing Plans in 2026

Choose the cheapest plan whose storage, operations, region and required features all fit. Starter and Builder have hard usage ceilings. Standard and Enterprise meter usage, with a monthly spending floor.

This is the database pricing schedule verified on October 5, 2026. A write unit is abbreviated WU and a read unit RU; neither means one vector or one query.

The paid usage rates vary by cloud and region. Standard includes a three-week trial with $300 in credits. Builder, Standard and Enterprise can use available AWS, Microsoft Azure and Google Cloud regions; Starter's database is restricted to AWS us-east-1. Source: Pinecone pricing.

The Standard minimum is applied to usage. If metered usage is $20, the bill is $50. If usage is $100, the bill is $100. You do not add another $50 to that $100. Enterprise works the same way with its $500 floor; Builder's $20 instead buys a capped allowance. Pinecone's cost documentation makes this distinction explicit.

Standard adds capabilities such as bulk import, backups and restore, role-based access control, which limits what different users can do, and SAML single sign-on. Enterprise adds a 99.95% uptime SLA, private endpoints, customer-managed encryption keys, audit logs, HIPAA compliance, Bring Your Own Cloud and Pro support. Buy Enterprise when a named requirement needs it: its published read and write rates are higher than Standard's, so a larger corpus alone does not make it the cheaper tier.

Object limits also matter. Starter allows up to five indexes and 100 namespaces per index, within one project and up to two users. Builder allows 10 indexes per project, 1,000 namespaces per index, five projects and up to five users. Standard allows 20 indexes per project, 100,000 namespaces per index and 20 projects; Enterprise allows 200, 1,000,000 and 100 respectively. Standard and Enterprise allow unlimited users. These are limits on organization structure, separate from the monthly operation budget.

When Starter and Builder Are Enough

Starter is sufficient for a small, stable corpus with modest retrieval traffic. Builder earns its $20 when you need more storage, more monthly units or a different cloud region and still fit its caps.

Pinecone Free Tier: Who Never Needs to Pay

The Pinecone free tier includes 2 GB of database storage, 2M monthly WUs and 1M monthly RUs, plus 1 GB of monthly egress, the data returned by reads. A personal knowledge base, small internal search feature or prototype can stay free indefinitely if its storage, reads, writes, returned data and region requirements stay within those allowances.

Free capacity is expressed in bytes, not a fixed vector count. With the example record used below, a 1,536-dimensional vector plus 1,000 bytes of metadata and an eight-byte ID, the storage allowance holds about 279,642 records. That is a storage-only ceiling; loading or querying them can exhaust a different allowance first.

A tiny namespace has a minimum query charge of 0.25 RUs. That makes 1M RUs enough for at most 4M minimum-sized queries, before fetches or other reads. A full 2 GB namespace uses 2 RUs per query, reducing the RU-only ceiling to 500,000 queries a month. Returning large payloads can hit egress before either number. Metering source.

Builder's $20 Covers Usage Until a Cap Stops It

Builder includes 10 GB, 5M monthly WUs, 2M monthly RUs and 10 GB monthly egress. Its database storage allowance is 10 GB; the separate Pinecone Assistant product has a different storage allowance.

At the stated record size, Builder's storage-only ceiling is about 1,398,210 vectors. One million records occupy 7.152 GB. At 100,000 whole-namespace queries a month, they consume 715,200 RUs, so the steady storage-and-read workload fits Builder for $20 if writes, egress and required features also fit.

The same corpus reaches Builder's RU ceiling at about 279,642 whole-namespace queries a month. One million such queries does not fit. Initial ingestion is another boundary: loading a million new records of roughly 7.15 KB each by upsert requires roughly 7.15M WUs in large batches, before request rounding, exceeding Builder's 5M monthly write allowance.

Monthly RU and WU limits apply across the organization. Pinecone's rate-limit docs also retain per-second limits on Standard and Enterprise: unlimited monthly usage does not promise unlimited request throughput.

Pinecone Vector Database Pricing: What a Query Costs

A query's read-unit cost depends on the namespace it targets. A namespace is a separate group of records inside an index. Think of it as the searchable room you enter: pricing follows the room's size even when you ask for only a few results.

Pinecone's current formula is 1 RU per GB of namespace size, with a 0.25 RU minimum per query. A 7.152 GB namespace therefore uses 7.152 RUs per query. A 715.2 GB namespace uses 715.2 RUs.

top_k, the number of results requested, does not change that RU formula. Nor do include_metadata and include_values. Those settings affect the response, and therefore can affect egress, but do not shrink the metered namespace.

For dense vectors without a sparse component, calculate stored bytes as:

Records × (ID bytes + metadata bytes + vector dimensions × 4 bytes).

A write unit measures write work. Upsert, which inserts or replaces records, uses 1 WU per KB of request with a five-WU minimum per request. Replacing a record also meters the existing record's size. Updates and deletes have their own write charges. This is why "one million writes" is an incomplete budget: you need bytes and operation type to translate writes into WUs.

For a RAG app, count database retrievals rather than chat messages. An illustrative 250,000 questions with four retrieval calls each produces one million database queries. An agent that searches repeatedly during a task can consume substantially more units than its user-facing interaction count suggests.

Pinecone Pricing Calculator: 1M, 10M and 100M Vectors

At one million database queries per month, the modeled Standard storage-and-read bills are $116.79, $1,167.92 and $11,679.22 for 1M, 10M and 100M vectors respectively, using the lowest published Standard read rate.

These are reproducible billing estimates, with the assumptions held constant:

• Dense vectors only, 1,536 dimensions per record.
• 1,000 bytes of average metadata and an eight-byte average ID, giving 7,152 bytes per record.
• Decimal GB, 1 GB = 1,000,000,000 bytes.
• The whole dataset sits in one namespace, and each query searches that namespace.
• 1,000,000 queries in a full month, with the dataset already loaded and its size unchanged.
• No ongoing writes, extra fetch/list calls, backups, imports, chargeable egress, inference, Assistant usage, support add-ons, discounts or taxes in this baseline.

Storage uses the live $0.33/GB/month rate. The primary calculation uses $16/M RUs; the sensitivity calculation uses $18/M RUs, the top of the published Standard range. These are endpoints of the published range, not a claim that every region has the lower price. Inputs: pricing and storage/read formulas.

For a uniform query scope, the monthly database estimate is:

Usage = stored GB × $0.33 + queries × max(0.25, namespace GB) ÷ 1,000,000 × RU price + WUs ÷ 1,000,000 × WU price.

Standard bill = max($50, total metered usage). Other billable products and add-ons must be handled according to their own pricing.

1M Vectors: $116.79 a Month at the Lower Read Rate

One million × 7,152 bytes gives 7.152 GB. Each query uses 7.152 RUs, so one million queries consumes 7.152M RUs.

Reads cost 7.152 × $16 = $114.432. Storage costs 7.152 × $0.33 = $2.36016. Together that is $116.79 a month after rounding, above Standard's $50 minimum. At $18/M RUs, the total becomes $131.10.

The lower-rate baseline works out to about $0.117 per 1,000 retrievals, including storage spread over that month's queries. Builder can hold this dataset, but cannot serve this whole-namespace query volume within its 2M-RU allowance.

Without writes or extras, this dataset grows into Standard's $50 minimum at roughly 416,316 monthly queries at $16/M RUs. Below that, its computed usage may be cheaper than $50, but its Standard invoice is still $50. Builder remains $20 while its own limits fit; there is no price crossover where Standard becomes cheaper for the same allowed workload.

10M Vectors: $1,167.92 a Month at the Lower Read Rate

Ten million records occupy 71.52 GB, producing 71.52 RUs per query and 71.52M monthly RUs at the same query volume.

Reads cost $1,144.32 and storage costs $23.60, giving $1,167.92 a month. At $18/M RUs, the total is $1,310.96. The lower-rate bill is about $1.168 per 1,000 retrievals after allocating storage.

The dataset already exceeds Builder's 10 GB storage cap. Reducing traffic alone cannot put this unchanged corpus back on Builder. You would need a smaller stored dataset as well as suitable read, write and egress volumes.

100M Vectors: $11,679.22 a Month at the Lower Read Rate

One hundred million records occupy 715.2 GB. Each whole-namespace query uses 715.2 RUs, so the monthly query workload consumes 715.2M RUs.

Reads cost $11,443.20 and storage costs $236.02, giving $11,679.22 a month. At $18/M RUs, the total becomes $13,109.62. The lower-rate baseline is about $11.679 per 1,000 retrievals after allocating storage.

Cheap storage does not make broad repeated retrieval cheap. At this scale, inspect the query boundary before accepting the bill: whether each customer needs the entire corpus is a budget decision as well as a data-model decision.

Add Writes and Enterprise at Their Own Rates

For Standard, each additional million WUs costs $4 to $4.50 at list price. Add that charge to usage before applying the minimum. Initial loading, re-embedding, which regenerates stored vectors, and ongoing memory updates belong in that write budget.

At Enterprise's $24 to $27/M RUs, the same no-write scenarios cost $500 at 1M vectors, because its minimum dominates; $1,740.08 to $1,954.64 at 10M; and $17,400.82 to $19,546.42 at 100M. Enterprise writes add $6 to $6.75 per million WUs. The higher bill buys its plan-specific controls and support, not a volume discount at published list rates.

For your own estimate, replace the record bytes and query scope before changing the vector count. Pinecone also provides an official pricing calculator for the selected cloud, region and workload.

The Same 100M Vectors Can Cost $350.45 to Query and Store

Searching only the relevant tenant's namespace changes this example from $11,679.22 to $350.45 a month. A tenant is one customer or account whose records can legitimately be searched separately.

Keep the same 100M records, but distribute them evenly across 100 tenants, each with 1M records in a 7.152 GB namespace. Keep the total monthly query count at one million across the application, and direct each request only to its tenant's namespace.

The reads now use 7.152M RUs for the month, costing $114.43 at $16/M RUs. Total storage remains 715.2 GB, costing $236.02. The resulting Standard storage-and-read bill is $350.45. Read-unit consumption falls 100-fold; the total bill falls by less because storage is unchanged.

A tenant filter inside one shared namespace does not get the same billing treatment. Pinecone's multitenancy documentation explicitly contrasts querying a tenant namespace with filtering a shared namespace: the latter still meters the shared namespace's size.

This saving requires a suitable search boundary. If every question must search all 100 namespaces, the fan-out recreates the aggregate read work. Keep global material searchable where the application needs it, and enforce the correct tenant scope through trusted application identity.

Costs Outside the Storage-and-Query Estimate

Your production budget must include ingestion, returned data and any products you enable. A vector count and monthly query count are sufficient for the narrow example above, not for every line of an invoice.

Initial ingestion and changing memory. Upserts consume WUs, even before the first user searches. Frequent overwrites meter the old and new record sizes. Batch small writes where practical to reduce the effect of the minimum request charge, then forecast ongoing updates separately from the initial load.

Response transfer. Starter includes 1 GB monthly egress and Builder 10 GB; reads can be blocked after those allowances. Standard and Enterprise include 100 GB/month, then $0.10/GB. Returning vectors or large metadata payloads can make a previously small response budget significant. Omitting vector values with include_values=false reduces response bytes; it does not reduce query RUs. Cost controls.

Import, backup and recovery. The published rates are $0.25/GB for bulk import, $0.10/GB/month for backups and $0.15/GB for restore, with additional cross-region transfer pricing where applicable. These are outside the worked storage-and-read baseline. For a large initial load, compare import with upsert rather than assuming every load must use the streaming write path.

Embeddings and reranking. An embedding converts content into a vector; reranking reorders retrieved candidates. Pinecone's paid embedding prices include $0.16/M tokens for llama-text-embed-v2 and $0.08/M tokens for multilingual-e5-large and pinecone-sparse-english-v0. The listed bge-reranker-v2-m3 and pinecone-rerank-v0 rate is $2/1,000 requests; cohere-rerank-4-fast uses $2/1,000 rerank units. These are separate meters from database RUs and WUs. Starter includes 5M embedding tokens per model monthly and Builder 10M; included bge-reranker-v2-m3 requests are 500 and 1,000 respectively. Current product rates.

Assistant versus database. Pinecone Assistant is a separate product for building chat and agent applications. Its paid storage is $3/GB/month, compared with the database's $0.33, and its input, output and context-processing token rates are $8, $15 and $5 per million respectively. Builder's 3 GB Assistant storage does not replace its 10 GB database storage. Reading the wrong product row produces the wrong budget.

Support and compliance. Builder and Standard include Free support; response SLAs can require a support add-on. Standard's optional HIPAA add-on is $190/month with a six-month minimum. Enterprise includes HIPAA compliance and Pro support. Keep these required fixed costs in the buying decision rather than expecting database units to pay for every service.

Monthly Billing Versus Annual Commitments

The annualized monthly floors are $240 for Builder, $600 for Standard and $6,000 for Enterprise. These are twelve months of the public monthly price or minimum, not discounted annual-plan quotes.

The live pricing page offers negotiated committed-use contracts rather than a universal annual discount percentage. Pinecone's docs describe $8,000 to $25,000 prepaid-credit purchases, with usage above the available credits billed at list price. Compare any offer with a representative usage forecast and read its expiry and unused-credit terms before making a commitment. Prepaid-credit documentation.

Promotional credits are a different category. Pinecone's subscription terms say they generally expire 30 days after issuance unless Pinecone specifies otherwise in writing. The Standard trial has its own stated 21-day duration; do not budget its $300 as recurring monthly credit.

The Standard Trial Ends at 21 Days or $300

The Standard trial blocks quotas when its time expires or its credits run out. It does not automatically put you back on the plan you used before the trial.

The trial documentation gives an organization $300 over 21 days. To restore access afterward, add a payment method for Standard, upgrade to Enterprise or downgrade to Starter. An organization starting from Builder is not charged its $20 Builder fee during the trial.

If you intend to return to Starter, plan the exit before loading production data: retained indexes must fit Starter's storage and region limits, and downgrading can delete resources you do not keep. Returning to Builder also requires its quotas and removal of unsupported features.

Each organization gets only one Standard trial. Organizations already on Standard or Enterprise, and organizations that initially subscribed through marketplace partners, cannot activate it. A trial is an evaluation budget, not a dependable production funding plan.

Pinecone Alternatives When a Cheaper Path Fits

An existing PostgreSQL database or a small free managed cluster can be cheaper when it already satisfies your workload. Compare the monthly commitment, the capacity it buys and the operating work you take on before treating a lower entry price as a replacement for the modeled Pinecone bill.

pgvector on Existing PostgreSQL or Supabase

pgvector is an open-source PostgreSQL extension for vector similarity search, with no separate paid software subscription. If you already run PostgreSQL and have suitable spare capacity, you may avoid a new database subscription; extra CPU, memory, disk, indexes, backups and administration still have a cost. Project source.

Supabase hosts PostgreSQL with the pgvector extension available. Supabase Pro starts at $25/month, with the first project included and 8 GB disk per project, then $0.125/GB. Compute upgrades can raise that bill.

Normalized to a year of monthly entry pricing, that is $300 for Supabase Pro versus $240 for Pinecone Builder and $600 for Pinecone Standard's minimum. Those are purchasing floors, not equivalent retrieval-capacity quotes. The 7.152 GB of logical records in the 1M-vector example is not a claim that the complete PostgreSQL database and its search indexes fit into 8 GB, or that base compute meets your application's requirements.

Qdrant Cloud for a Small Free Prototype

Qdrant is a vector database with a managed Cloud offering. Qdrant Cloud Free is free forever for a single-node cluster with 0.5 vCPU, 1 GB RAM and 4 GB disk. Its published fit is testing and prototypes, without high availability.

That is a $0 subscription floor, but the 7.152 GB logical corpus in this article's smallest worked example already exceeds the free cluster's disk allowance before indexing overhead. Qdrant's paid Standard tier charges for resources, so quote the necessary cluster rather than carrying its free price into a million-vector production comparison.

Use the best vector database comparison to choose among these routes when PostgreSQL reuse, a dedicated vector service or managed operations could change the purchase. A cheaper path fits when its capacity, recovery requirements and engineering work fit too.

Pinecone Self-Hosted: BYOC and Local Are Different

Pinecone self-hosted is not a free production license you can infer from its Docker tooling. Bring Your Own Cloud runs managed Pinecone in your cloud account, requires Enterprise and has custom pricing. Include your infrastructure as well as the commercial offer when comparing it.

Pinecone Local is an in-memory Docker emulator that avoids Pinecone usage and storage fees for local development. It is unsuitable for production, and its records do not persist when it stops. It can reduce development spend; it cannot replace the production database budget.

Forecast Your Bill Before Launch

Measure the scope and units of a representative workload, then apply the selected region's prices. Treat the article's examples as a calculation method you can replace with your own inputs.

1. 1

   Record the stored bytes and search boundary

   Calculate average vector, ID and metadata bytes. Identify which namespace each retrieval must search, including any calls that fan out across multiple namespaces.

2. 2

   Count database calls and mutations

   Separate user questions from retrieval calls. Include agent loops, fetch/list requests, initial loading, ongoing upserts, overwrites and deletes. Capture the RUs returned by representative reads and track write-unit usage.

3. 3

   Price the month and apply the floor

   Multiply each usage meter by its rate. Add chargeable egress, backups and enabled products, then account for the plan minimum and any separately billed add-ons. For Starter or Builder, compare every meter with its cap instead.

4. 4

   Check the billing report

   On Standard or Enterprise, an organization owner or billing admin can open Settings > Usage, inspect costs by SKU, project or service, and download a CSV. Cost data can lag actual usage by up to three days, so use returned units for immediate feedback as well as the billing report.

Pinecone's usage-monitoring guide documents those controls. For sustained high read throughput, also ask for a dedicated-read-node estimate; compare its provisioned capacity with your actual traffic before assuming on-demand remains the cheapest mode.

Frequently Asked Questions

Is Pinecone free to use?

Yes. Starter is $0 within 2 GB database storage, 2M WUs and 1M RUs per month, plus its other caps, and its database region is AWS us-east-1. Builder costs $20 flat, Standard has a $50 monthly minimum and Enterprise has a $500 minimum. The Standard trial is temporary credit rather than a permanent free production allowance.

How much do you make on Pinecone research?

These prices cover Pinecone's vector database and cannot establish survey earnings. They describe what a developer pays for infrastructure.

Is Pinecone a good vector database?

It is a sensible candidate when a managed service and its deployment controls justify the projected bill. Compare scoped usage costs and required features, then evaluate retrieval quality with representative application data before choosing.

How much does vector database cost?

It depends on stored bytes, query scope, traffic, mutations and deployment. Under this article's stated assumptions, Pinecone Standard storage and reads cost $116.79 for 1M vectors, $1,167.92 for 10M and $11,679.22 for 100M at one million monthly whole-namespace queries and the lowest listed read rate.

What is replacing vector databases?

A vector extension in an existing database can replace a separate vector service for suitable workloads. pgvector adds vector similarity search to PostgreSQL; whether it fits depends on your data model, capacity and operating requirements.

Which is the cheapest database?

For an application already on PostgreSQL with spare capacity, pgvector can avoid a new service subscription. For a small managed prototype, Pinecone Starter and Qdrant Cloud Free have $0 entry prices with different limits. Compare the complete deployed workload before naming a cheapest production option.

Which database is completely free?

Free database software and free hosted infrastructure are different things. pgvector has no separate software subscription, but hosting can cost money. Pinecone Starter and Qdrant Cloud Free provide capped hosted capacity.

What is the best database right now?

For this buying decision, choose by searchable scope, permissions, where the source data already lives and who will operate the service. Pinecone fits a managed-service preference; pgvector fits PostgreSQL reuse. Neither answer follows from vector count alone.

What is the easiest database to use?

A managed service such as Pinecone removes cluster administration from your workload. If your product already uses PostgreSQL, keeping vector search there may reduce the number of services you integrate. Choose the route that removes the work your application would otherwise need.

Does Pinecone offer a student discount?

No student-specific database discount is advertised on the live pricing page checked for this article. Starter remains the public free option; ask Pinecone about any eligibility-specific offer before including it in a budget.

Does Pinecone offer refunds?

Pinecone's subscription terms do not provide a general refund right for ordinary cancellation. They describe specific exceptions for unused prepaid amounts after qualifying termination. Check the applicable order and the current subscription terms before prepaying.

Has Pinecone pricing changed?

Yes. Pinecone's May 2026 launch announcement introduced Builder at $20/month, with availability on May 6. A budget that assumes your only choices are free Starter or Standard's $50 minimum misses that current middle tier.

Get the AI Business Workflow Audit Checklist

Map the data, permissions, handoffs and operating costs before adding another service to your AI workflow.