DynamoDB vs Cosmos DB

Quick verdict

Amazon DynamoDB is AWS's serverless key-value and document database. You create tables, not servers: each item is addressed by a partition key and optional sort key, and is read with GetItem, Query and Scan or a subset of PartiQL. Tables use on-demand (pay per request) or provisioned capacity, global tables replicate across Regions, and since 5 August 2026 vector indexes and the SearchVectors API add similarity search on on-demand tables. See our Amazon DynamoDB profile.

Azure Cosmos DB is Microsoft's fully managed NoSQL and vector database. Each account uses one API: the native API for NoSQL (JSON documents queried with a SQL-like language) or compatibility APIs for MongoDB, Apache Cassandra, Apache Gremlin and Table. Items are partitioned by a partition key, every operation is costed in request units (RUs), and you can add Azure regions to an account at any time. See our Azure Cosmos DB guide for APIs, capacity modes and worked price examples.

Side by side

Key differences

Data model, APIs and querying

DynamoDB has one data model and API. You design keys and secondary indexes around known questions, often with several entity types sharing one table, and anything that does not match a key or index becomes a Scan. PartiQL gives SQL-like syntax for single-table statements but no joins.

Cosmos DB asks you to choose an API per account. The native API for NoSQL indexes every property by default and supports a SQL-like language with filters, projections, aggregates and JOIN within an item (up to 10 per query by default). The MongoDB, Cassandra, Gremlin and Table APIs let existing drivers connect, with documented feature gaps; see MongoDB Atlas vs Azure Cosmos DB for the MongoDB case.

-- DynamoDB PartiQL: one customer's orders by key
SELECT * FROM "AppTable"
WHERE PK = 'CUSTOMER#123' AND begins_with(SK, 'ORDER#');

-- Cosmos DB API for NoSQL: the same within one partition
SELECT c.id, c.total, c.orderDate FROM c
WHERE c.customerId = "123" AND c.type = "order"
ORDER BY c.orderDate DESC

Capacity and pricing units

DynamoDB counts work in request units sized by data: a write request unit covers up to 1 KB, a strongly consistent read unit up to 4 KB, an eventually consistent read costs half and a transactional read or write costs double. On-demand bills per million units; provisioned bills per capacity unit per hour, with auto scaling and reserved capacity.

Cosmos DB normalises every operation into RUs: Microsoft's reference is that a point read of a 1 KB item by ID and partition key costs 1 RU, while writes, queries, larger items, more indexed properties and stronger consistency cost more (strong and bounded staleness reads cost about twice as much). You provision RU/s in steps of 100 (manual, minimum 400 RU/s per container), set an autoscale maximum (scaling between 10% and 100% and billed at the hourly peak), or use serverless, which is limited to one region. In both services, item size and access pattern drive cost more than data volume, so estimate with your own operations.

Consistency and transactions

DynamoDB offers two read modes: eventually consistent (default) and strongly consistent, which is not available on global secondary indexes. TransactWriteItems and TransactGetItems group up to 100 actions on distinct items, across tables, within one account and Region.

Cosmos DB offers five consistency levels, with session the most widely used according to Microsoft. Its ACID transactions use snapshot isolation but are scoped to one logical partition key in a container: a transactional batch holds up to 100 point operations, up to 2 MB, and must finish within 5 seconds; stored procedures written in JavaScript give the same scope. If you need atomic updates across unrelated entities, DynamoDB's cross-table transactions are the more flexible model; in Cosmos DB you design the partition key so that related items share it.

Multi-region replication

DynamoDB global tables replicate a table across Regions with every replica accepting writes. The default, multi-Region eventual consistency (MREC), resolves conflicts with last writer wins. Multi-Region strong consistency (MRSC) replicates synchronously across exactly three Regions (three replicas, or two plus a witness), in a listed set of Regions, and does not support transactions, TTL or local secondary indexes; the mode is fixed at creation.

Cosmos DB lets you add or remove any Azure region on an account and enable multiple write regions. Conflicts are resolved by last writer wins on the system timestamp (or, in the API for NoSQL, a custom numeric property) or by a custom merge stored procedure, which can be set only when the container is created. Microsoft states that accounts with multiple write regions cannot use strong consistency; strong consistency is available with a single write region, where a write commits only after every region acknowledges it. Data and throughput are billed in every region you add.

Vector search for AI applications

DynamoDB added vector indexes on 5 August 2026. You define them with CreateTable or UpdateTable and query with SearchVectors, which performs approximate nearest neighbour search with cosine, dot product or Euclidean distance. Vector indexes require on-demand capacity (5 per table by default, adjustable to 20), are indexed asynchronously after each write, work with global tables and are not supported by DAX.

Cosmos DB for NoSQL stores vectors inside documents and offers flat (exact, up to 505 dimensions), quantizedFlat and DiskANN indexes (up to 4,096 dimensions), queried with the VectorDistance function alongside normal WHERE filters. Microsoft notes that vector search is not supported on accounts with shared database throughput and cannot be disabled on a container once enabled.

Pricing and licensing

Amazon DynamoDB (US East, N. Virginia, Standard table class, USD, from the AWS DynamoDB pricing pages on 7 November 2026): on-demand 0.625 per million write request units and 0.125 per million read request units; storage 0.25 per GB-month; provisioned 0.00065 per WCU-hour and 0.00013 per RCU-hour; point-in-time recovery 0.20 per GB-month; global table replicated writes 0.625 per million. The Standard-IA class lowers storage to 0.10 per GB-month at higher request prices. Vector search is billed separately by data written to vector indexes, data searched and vector index storage; check the on-demand pricing page for the rates. The free tier gives 25 GB and 25 provisioned WCUs and RCUs per Region each month.

Azure Cosmos DB (East US, USD, from Microsoft's Azure Retail Prices API on 7 November 2026, request unit accounts): standard provisioned throughput 0.008 per 100 RU/s per hour (about 5.84 a month); multi-region write throughput 0.016 per 100 RU/s per hour; autoscale 0.012 per 100 RU/s per hour, billed at the highest RU/s reached each hour; serverless 0.25 per million RUs; transactional storage 0.25 per GB-month in each region. The lifetime free tier covers 1,000 RU/s and 25 GB for one account per subscription.

Putting them side by side. The per-unit prices are not comparable directly, because a DynamoDB request unit and a Cosmos DB RU measure different amounts of work. As arithmetic only: 400 RU/s provisioned in one Cosmos DB region with 10 GB stored is 4 x 5.84 + 10 x 0.25 = USD 25.86 a month; on DynamoDB on-demand, 10 million 1 KB writes and 50 million 4 KB strongly consistent reads with 10 GB stored is 10 x 0.625 + 50 x 0.125 + 10 x 0.25 = USD 15.00 a month before the free tier. These workloads are not equivalent; use the AWS Pricing Calculator and the Azure pricing calculator with your own traffic.

Pricing checked on the vendors' official pages on 7 November 2026. Prices change; confirm before buying.

Where each one leads

Limitations

When to choose each

When neither is right

• You need to run the same database on more than one cloud or on premises: see DynamoDB vs MongoDB and Cassandra vs ScyllaDB.
• You want MongoDB itself as a managed service: see MongoDB Atlas vs Azure Cosmos DB.
• Your data is relational and needs joins and multi-row transactions: see DynamoDB vs PostgreSQL or distributed SQL in CockroachDB vs YugabyteDB.
• You need analytics over operational data: both vendors point analytical workloads to other services.

Final recommendation

Frequently asked questions

Sources

• Amazon DynamoDB on-demand pricing
• DynamoDB Developer Guide: How global tables work
• DynamoDB Developer Guide: Using vector indexes
• AWS What's New: DynamoDB vector search (August 2026)
• Microsoft Learn: Azure Cosmos DB service quotas and limits
• Microsoft Learn: Transactional batch operations
• Microsoft Learn: Consistency levels
• Microsoft Learn: Conflict resolution policies
• Microsoft Learn: Integrated vector store in Azure Cosmos DB
• Azure Cosmos DB pricing: standard provisioned throughput

Checked November 2026.

How we research comparisons: our editorial method.