S3 Cost Optimization: Storage Classes, Traps & Levers

Amazon S3 is the cheapest storage most companies own and one of the most wasteful lines on their bill, and both statements are true at once. The base rate is genuinely low: about two cents per gigabyte per month. The waste comes from everything around that rate. Data sits in the expensive tier years after anyone last read it. Old object versions pile up invisibly behind the current ones. Incomplete uploads occupy space nobody knows exists. And a single terabyte served to the internet costs four times what it costs to store.

None of that is a flaw in Amazon S3; it is the predictable result of a service designed to accept anything you put in it and keep it forever unless told otherwise. S3 cost optimization is the practice of telling it otherwise. This guide covers what each storage class actually costs in 2026, the charges that do not appear on the storage line, and eight levers that reliably cut the bill, most of which are a single lifecycle rule or configuration change.

Key Takeaways

• S3 storage classes span a 23x price range, from Standard at $0.023 per GB-month to Glacier Deep Archive at $0.00099, so storage class is the biggest lever.
• Egress is the charge that surprises people: serving 1 TB to the internet costs about $90, roughly four times the cost of storing it.
• Three kinds of silent waste accumulate in most buckets: non-current versions, incomplete multipart uploads, and data that has not been read in months but sits in Standard.
• Lifecycle policies and Intelligent-Tiering automate the largest savings; Storage Lens shows you where to point them.
• Requests, retrieval fees, and minimum storage durations mean the cheapest class is not always cheapest for your access pattern.

How S3 Pricing Works

S3 bills on four dimensions, and optimizing only the first is why most efforts stall. The full rate card is on the official Amazon S3 pricing page, and the four dimensions are storage (per GB-month, varying by storage class), requests (per thousand, with PUT, COPY, POST, and LIST costing about ten times more than GET), data transfer (free in, free within a region, billed out to the internet and between regions), and features such as replication, Storage Lens advanced metrics, and management tools. Storage is usually the largest line, but egress and requests are the ones that produce surprises, because they scale with how data is used rather than how much of it exists.

S3 Storage Classes and What They Cost

AWS describes each tier on its S3 storage classes page; the table below pulls the 2026 rates together for US East (N. Virginia). Prices vary by region, so verify yours before modeling a budget.

Two details in that table decide whether a move saves money. Retrieval fees mean that data you read often should stay in Standard even though it is the most expensive per gigabyte; a terabyte read twice a month from Standard-IA costs more than leaving it in Standard. And minimum durations mean deleting an object early bills the full period, so lifecycle rules should transition data only once you are confident it will stay put.

The Hidden Costs That Inflate S3 Bills

Most S3 overspend is not on the storage line at all. Five charges account for the bulk of it.

• Internet egress. Data out to the internet costs about $0.09 per GB for the first 10 TB. Storing 1 TB costs about $23 a month; serving it to users once costs about $90. Routing through CloudFront makes the S3-to-CloudFront leg free.
• Non-current versions. With versioning enabled, every overwrite keeps the old copy. Without an expiration rule, versions accumulate until they can double the storage bill while remaining invisible in the console's default view.
• Incomplete multipart uploads. Large uploads that fail or are abandoned leave their parts stored and billed indefinitely. Audits routinely find 5 to 15 percent of S3 spend in these fragments.
• NAT Gateway processing. EC2 instances in private subnets reaching S3 through a NAT Gateway pay about $0.045 per GB processed. A gateway VPC endpoint for S3 is free and removes the charge entirely.
• Request volume and Glacier metadata. LIST and PUT requests cost roughly ten times GET; code that paginates a huge bucket to find one key can fire thousands of billable calls. Glacier classes also store about 40 KB of metadata per object, which penalizes archiving millions of tiny files.

These are the S3-specific versions of the general pattern we describe in our guide to where cloud waste hides: none is a bug, all are the residue of teams moving fast, and each is cheap to fix once seen.

Eight Levers That Reduce S3 Costs

1. Find out what you actually have

Start with visibility, because optimizing blind moves data you need and leaves data you do not. S3 Storage Lens gives an organization-wide view of storage by class, bucket, and age, and its free metrics show which buckets hold data untouched for 90 days or more. That report is your target list for everything below.

2. Set lifecycle policies for data with known patterns

Where you know how data ages, a lifecycle configuration transitions objects down the tiers on a schedule: logs to Standard-IA at 30 days and Glacier Instant at 90, compliance records straight to Deep Archive at 180, temporary exports expired at 7. This is usually the single largest saving available, and it is one rule per bucket.

3. Use Intelligent-Tiering for data with unknown patterns

Where access is unpredictable, S3 Intelligent-Tiering moves each object between tiers automatically based on its own access history, for a monitoring fee of $0.0025 per 1,000 objects, with no retrieval charges. Enable the optional Archive and Deep Archive access tiers for buckets that hold genuinely cold data. The one caution: objects under 128 KB never leave the frequent tier, and the per-object fee makes it a poor fit for buckets full of millions of tiny files.

4. Expire non-current versions

If versioning is on, add a lifecycle rule that transitions non-current versions to a cheaper class after a short window and expires them after a defined retention period. Keep enough history to recover from mistakes, typically 30 to 90 days, and let the rest go. On long-lived versioned buckets this alone can halve the storage line.

5. Abort incomplete multipart uploads

Add a lifecycle rule to abort incomplete multipart uploads after seven days. It is a single setting, it recovers the 5 to 15 percent of spend those fragments typically consume, and there is no downside, since a genuinely in-progress upload finishes long before the window closes.

6. Stop paying for egress you can avoid

Put CloudFront in front of buckets that serve public assets, so the S3-to-CloudFront transfer is free and edge caching cuts origin requests. Replace NAT Gateway paths to S3 with a free gateway VPC endpoint. Keep compute and the buckets it reads in the same region, since cross-region transfer bills about $0.02 per GB and cross-region replication pays transfer, storage, and requests twice.

7. Reduce request costs

Batch small writes into larger objects where the application allows, avoid repeated LIST calls by tracking keys in a database or using S3 Inventory, and use S3 Select or Athena to read only the bytes you need rather than pulling whole objects. Request costs rarely dominate, but on high-object-count workloads they can rival storage.

8. Match the class to the access pattern, not the price list

The cheapest class per gigabyte is not the cheapest for every workload. Model retrieval fees and minimum durations against your real read frequency before moving anything. Frequently read data belongs in Standard; data read a few times a year belongs in Glacier Instant; data that must exist but will almost never be read belongs in Deep Archive. The wrong move in either direction costs money.

A Worked Example

Consider 100 TB of log and compliance data that has been sitting in S3 Standard because that is where it landed. At the tiered Standard rate, that costs roughly $2,250 a month, about $27,000 a year.

Suppose Storage Lens shows that 20 TB is read regularly, 30 TB is read occasionally, and 50 TB has not been touched in a year and is kept for compliance. Leaving the 20 TB in Standard costs about $460. Moving the 30 TB to Glacier Instant Retrieval costs about $120. Moving the 50 TB to Deep Archive costs about $50. The new total is roughly $630 a month, a saving of about $1,600 a month or $19,000 a year, from two lifecycle rules. Add an incomplete-upload abort rule and version expiration and most estates find another 10 to 20 percent on top.

At larger scale the same logic compounds; our case study on petabyte-scale S3 cost optimization walks through the same levers at a thousand times the volume.

When Moving Data Costs More Than It Saves

Three honest caveats keep S3 optimization from backfiring.

• Retrieval-heavy data. If you read an object more than once or twice a month, the retrieval fees on IA and Glacier classes can exceed the storage saving. Keep it in Standard or Intelligent-Tiering.
• Millions of tiny objects. Per-object monitoring fees, request charges, and Glacier's 40 KB metadata overhead can make archiving small files more expensive than leaving them. Aggregate first, then archive.
• Uncertain retention. Minimum durations of 30, 90, or 180 days bill in full if you delete early, so do not transition data you might remove next month.

The Bottom Line

S3 cost optimization is unusually forgiving work: the levers are documented, most are a single rule, and the savings are large because the price gap between tiers is so wide. Find out what you have with Storage Lens, move known-pattern data with lifecycle rules and unknown-pattern data with Intelligent-Tiering, expire old versions, abort dead uploads, and stop paying for egress that a CDN or a VPC endpoint would make free. Then treat it as ongoing hygiene rather than a project, because new data lands in Standard every day. The same discipline applies across the AWS bill, which our guides to AWS cost optimization best practices and understanding AWS pricing across EC2, S3, EBS, and RDS cover in full.