S3 Storage Classes and Lifecycle

Topics Covered

How S3 Stores and Charges

Buckets, objects and keys

Durability and availability

The four parts of an S3 bill

Why requests can dominate

Data transfer out

Versioning and its hidden storage

Incomplete multipart uploads

Seeing where the money goes

Prorating and averaging

Requests you did not make directly

A worked example

Free and nearly free operations

Pricing differs by region

S3 is the most widely used service on AWS and one of the most misunderstood on the bill. It looks like a single product with a single price, a few cents per gigabyte per month, and in practice it is a family of storage classes with different prices for storage, requests, retrieval and minimum durations, plus charges for data transfer, replication, inventory and monitoring. Teams that understand how S3 charges can cut its cost by half or more without losing anything they need; teams that do not often discover that their largest S3 line items are not storage at all.

Buckets, objects and keys

A bucket is a container in one region with a globally unique name. An object is a blob of up to five terabytes with a key, metadata and optional tags. Keys look like paths, logs/2026/11/11/app.gz, but S3 has no directories: the slashes are just characters, and the console's folders are a display convention over key prefixes. Prefixes matter for two practical reasons: lifecycle rules and permissions can target them, and request rate limits apply per prefix, at least 3,500 write and 5,500 read requests per second per prefix, with S3 scaling partitions automatically as traffic grows.

Durability and availability

S3 Standard is designed for eleven nines of durability, storing data redundantly across at least three Availability Zones, and for 99.99 percent availability. Durability is about losing data, availability about being able to reach it at a given moment. The two are easy to confuse: an availability event makes objects temporarily unreachable without losing them, and data loss in S3 almost always comes from people and programs deleting or overwriting objects, not from hardware. Versioning, Object Lock and replication protect against that far more than any storage class choice.

The four parts of an S3 bill

Every storage class charges in up to four ways. Storage is billed per gigabyte month, prorated by the hour. Requests are billed per thousand, with writes, PUT, COPY, POST and LIST, costing more than reads, GET and SELECT. Retrieval is billed per gigabyte for the infrequent access and archive classes. Data transfer out to the internet, and between regions, is billed per gigabyte, while transfer in and transfer to services in the same region is free. On top of these sit features with their own charges: replication, S3 Inventory, Storage Lens advanced metrics, Intelligent Tiering's per object monitoring fee and Object Lambda.

The archive bucket paid almost only for storage. The thumbnail bucket paid mostly for requests, and the public download bucket mostly for data transfer out.

Why requests can dominate

Small objects make requests expensive relative to storage. A bucket of a hundred million one kilobyte objects holds only about a hundred gigabytes, costing a couple of dollars a month to store, but writing them costs hundreds of dollars and listing or reading them repeatedly costs more. Workloads that write many tiny objects, such as per event JSON files from a stream, telemetry fragments or thumbnails, often have request charges many times their storage charges. Batching small records into larger objects, with formats such as Parquet or compressed JSON lines, is the usual fix, and it also speeds up anything that later reads them.

Data transfer out

Serving objects directly from S3 to the internet charges data transfer out per gigabyte, with tiered discounts at high volume. For public downloads, images and video, this is often the largest S3 related cost. CloudFront in front of S3 reduces it, because transfer from S3 to CloudFront is free, CloudFront's own transfer prices are lower, and cached objects avoid repeated reads from S3 altogether. For large public datasets, Requester Pays buckets shift transfer and request charges to the people downloading.

Versioning and its hidden storage

Versioning keeps every version of every object, so overwrites and deletes become recoverable. A delete in a versioned bucket adds a delete marker and keeps the previous version, which is still billed. Buckets with versioning and frequent overwrites, such as those holding build artefacts under fixed names or files rewritten nightly, can accumulate noncurrent versions that cost several times the current data. Lifecycle rules that expire noncurrent versions after a set number of days, or keep only a set number of newer versions, keep versioning's protection without unbounded growth.

Incomplete multipart uploads

Large objects are uploaded in parts with multipart upload, and the parts of an upload that is never completed or aborted remain stored and billed, invisible in normal object listings. A pipeline that crashes during uploads can leave terabytes of orphaned parts. A lifecycle rule that aborts incomplete multipart uploads after a few days belongs on every bucket, and Storage Lens reports how many bytes are held in them.

Seeing where the money goes

Cost Explorer breaks S3 charges down by usage type, which separates storage by class, requests by type, retrieval and transfer. S3 Storage Lens shows storage and activity by account, region, bucket and prefix, including noncurrent versions and incomplete uploads, and its advanced tier adds activity metrics and recommendations. S3 Inventory produces daily or weekly lists of every object with its size, class, last modified date and encryption status, which can be queried with Athena to answer questions such as how much data has not been modified in a year. These three tools together answer almost every question about where S3 money goes.

Prorating and averaging

Storage is metered as byte hours and billed as the average for the month, so a terabyte stored for half a month costs half a terabyte month. A short lived spike of data, such as a temporary export kept for a day, therefore costs very little in Standard, while the same export copied into a class with a minimum duration costs the full minimum. It also means the storage line reflects the whole month's average, so cleanup work shows its full effect only in the following month, which matters when reporting savings.

Requests you did not make directly

Several features generate requests on your behalf. Lifecycle transitions are billed as requests per object moved. Replication issues writes at the destination for every replicated object. S3 Inventory and Storage Lens have their own per object charges. Event notifications are free from S3, but the Lambda, SQS or EventBridge targets they trigger are not. Even a well meaning backup tool that lists a bucket nightly to find changes can run up list charges larger than the bucket's storage. When request charges look higher than the application's traffic explains, CloudTrail data events or server access logs show which principal is making them.

A worked example

Consider three buckets with similar headline sizes. An archive bucket of 50 terabytes of database dumps, each several gigabytes, written weekly and almost never read, pays nearly all of its bill in storage, and is the bucket where storage class choices matter most. A thumbnail bucket of 2 terabytes holding 400 million small images, written continuously and read by an application without a cache, pays more in requests than in storage, and storage classes barely matter while a cache or larger objects would. A public download bucket of 5 terabytes of installers, served directly to users, pays mostly for data transfer out, and the fix is CloudFront, not lifecycle. The same advice applied to all three would help one and harm the others, which is why the bill should be read by usage type before any change.

Free and nearly free operations

Some operations cost nothing: data transfer in, transfer between S3 and other services in the same region, DELETE requests, and lifecycle expirations. Others are cheap enough to ignore at normal volumes. Knowing what is free changes designs: emptying a prefix with an expiration rule costs nothing, while copying it elsewhere first costs requests and storage for the copy.

Pricing differs by region

S3 prices vary by region, with US East and US West Oregon usually cheapest and some regions costing a third more for storage and much more for transfer. Data placed in a region for latency or residency reasons carries that region's prices, and buckets read by compute in a different region add transfer charges on every read. Keeping data in the same region as the compute that reads it is one of the simplest cost rules, and one of the most often broken by analytics jobs pointed at a bucket created years earlier somewhere else.

Storage is only one of four charges

S3 bills storage, requests, retrieval and data transfer, and for many buckets storage is not the biggest. Small objects make requests expensive, public downloads make transfer expensive, and versioning and incomplete multipart uploads make storage silently grow. Read the bill by usage type before choosing storage classes.