Amazon S3 pricing: the complete 2026 guide

S3 storage costs are the AWS line item that finance leaders consistently underestimate. The storage charge itself is straightforward. Everything around it, data transfer, request volume, retrieval fees, and now AI-specific storage types, is where the budget surprises live.

If your company is building AI products, S3 is almost certainly inside your AI spend. Training datasets, model checkpoints, vector embeddings, feature stores, inference logs, all of it goes through S3, and all of it is billed across multiple dimensions that a single line in Cost Explorer won’t show you.

A September 2025 Crayon survey found that 94% of IT leaders are still struggling to optimize cloud costs. S3 is rarely the only culprit, but it’s almost always part of the picture. This guide covers every billing dimension and explains the new AI-era storage types (S3 Vectors, S3 Tables, S3 Metadata).

At CloudZero, we help finance and engineering teams at companies like Duolingo, Grammarly, Skyscanner, and PicPay break their S3 costs down by team, product, and AI workload. The patterns we see most; the data in the wrong storage class, unallocated AI spend, forgotten replication rules; they’re the same ones this guide covers.

Jump to: AWS Data Transfer Costs | AI Costs in 2026 | ChatGPT Pricing | Amazon Bedrock Pricing

How does Amazon S3 pricing work?

S3 charges across eight independent billing dimensions simultaneously. Most teams monitor one, storage, and get surprised by the others. Finance leaders reviewing an AWS bill need to know which meter generated which charge, because they don’t all roll up neatly.

Here are the Amazon S3 pricing dimensions:

Before we get into each: the AWS free tier changed July 15, 2025. New AWS accounts no longer receive the old ‘5 GB free + 12-month trial’. Instead they get $200 in AWS credits valid for 6 months, applicable across all services. Accounts created before July 15, 2025 keep the original free tier terms. 

For finance teams: the free tier change means new accounts can’t rely on a fixed 5 GB free storage baseline for budget modeling. The $200 credit runs out in 6 months and doesn’t renew. Plan accordingly.

What are the 8 S3 storage classes and how do they compare?

Storage class selection is the single largest controllable variable in your S3 bill. The difference between the most expensive and cheapest class is 23x. For a finance team looking at a $500,000/year S3 bill, getting this right is worth more than any other optimization, and it usually takes one afternoon and a Lifecycle policy.

All prices below are for us-east-1. Other regions cost 10–30% more. 

S3 Standard: right for active data, expensive for everything else

S3 Standard is $0.023 per GB per month for the first 50 TB, $0.022/GB for the next 450 TB, and $0.021/GB above 500 TB. Volume discounts are real but only kick in at scale most teams don’t reach.

The finance risk: S3 buckets accumulate data quietly, and without Storage Lens or Lifecycle policies, yesterday’s active ML training dataset becomes this month’s cold data paying hot storage rates. Cloud cost management tools surface this pattern automatically across multi-account environments, which is usually where it’s hardest to spot.

S3 Standard-IA: 46% cheaper, but read the fine print

Standard-IA costs $0.0125 per GB per month, 46% cheaper than Standard, but adds a per-GB retrieval fee on top. Millisecond retrieval, multi-AZ durability, same 99.999999999% (eleven nines) of object durability as Standard. Good for backups, DR copies, logs queried occasionally. 

The 30-day minimum storage period is the billing trap. Delete an object at day 15 and you’re still billed for 30. Short-lived data belongs in Standard, the savings don’t materialize until you’ve held data for the full minimum period.

Comparing GCP storage pricing? GCP Nearline follows a similar access-fee model but prices differently.

S3 One Zone-IA: 20% cheaper, one AZ instead of three

One Zone-IA stores data in a single Availability Zone, that’s how AWS gets to $0.01 per GB per month, 20% cheaper than Standard-IA. If that AZ goes down, you lose access until it recovers. 

Appropriate for reproducible data: thumbnail images, cached API responses, processed versions of originals held elsewhere in a more resilient class. Not appropriate for anything you couldn’t recreate. That’s not a judgment call, it’s what the 99.5% availability SLA says.

S3 Glacier Instant Retrieval: archive pricing, no patience required

Glacier Instant costs $0.004 per GB per month, 68% cheaper than Standard-IA, with millisecond retrieval. AWS’s intended use case: data you access once a quarter, compliance records, medical images, quarterly financial reports, audit logs.

The 90-day minimum storage period is the key budget constraint. Bill an object for 90 days even if you delete it at day 30. For AI workloads, this is the right class for archived training data that you might occasionally need to reproduce an experiment, fast access when you need it, minimal cost when you don’t.

S3 Glacier Flexible Retrieval: long-term archiving, hours not milliseconds

Glacier Flexible (formerly plain ‘S3 Glacier’) costs $0.0036 per GB per month. Retrievals take minutes to 12 hours depending on tier. Standard Bulk retrievals are free.

Best for: compliance archives you’ll never touch except in a legal hold, old database backups, data that must be retained for years. Aurora vs RDS transaction logs from either are natural candidates, keep them years, retrieve them almost never.

S3 Glacier Deep Archive: the cheapest cloud storage AWS sells

Glacier Deep Archive starts at $0.00099 per GB per month. Under $1 per TB per month. About $1,000 to store a petabyte for a year. Retrieval takes up to 12 hours. 180-day minimum storage period.

The comparison is stark: storing 100 TB of cold compliance data in S3 Standard costs $2,305/month. The same 100 TB in Glacier Deep Archive costs $101.38/month. That’s $26,944/year from a single Lifecycle policy. Most finance teams would consider that a meaningful line item.

S3 Intelligent-Tiering: the right default when access patterns are unknown

Intelligent-Tiering monitors per-object access and automatically moves data between five tiers: Frequent Access ($0.023/GB), Infrequent Access ($0.0125/GB after 30 days), Archive Instant ($0.004/GB after 90 days), plus two opt-in deep archive tiers. No retrieval fees. No transition charges.

The cost: $0.0025 per 1,000 objects per month monitoring fee for objects over 128 KB. Pays for itself on large buckets with unpredictable or seasonal access. For AI workloads with variable training cycles, data that’s hot during a sprint and cold between them, Intelligent-Tiering often outperforms any manually-managed Lifecycle policy.

New from 2025: S3 Tables now supports Intelligent-Tiering, meaning analytics workloads on Iceberg tables can also benefit from automatic tiering.

S3 Express One Zone: ML training performance, at a price

Express One Zone costs ~$0.11 per GB per month, 4.8x more than Standard, and delivers consistent single-digit millisecond latency at hundreds of thousands of requests per second. After April 2025 price cuts (85% on GET, 55% on PUT, 31% on storage), it became viable for high-value ML training workloads.

The ROI case: SageMaker training jobs reading from Express One Zone-colocated storage see 40–60% shorter training times from eliminated I/O wait. Athena queries run 2–3x faster. At $50,000/month in GPU compute, a 50% reduction in training time is worth far more than the storage premium.

The boundary condition: this only pencils out when I/O wait is the actual bottleneck. For ECS vs EC2 architectures doing standard data processing, it usually doesn’t. Model the compute savings before paying the storage premium.

How do S3 data transfer costs work?

Data transfer is the billing dimension finance teams underestimate most. Storage costs are predictable and visible. Transfer costs are neither, they depend on where data is going, what’s routing it, and decisions that were made by an engineering team months ago.

The transfer rule that saves most teams money: data INTO S3 is always free. The first 100 GB per month out to the internet is free across all AWS services combined. Everything above that costs money, and the rate depends on the destination.

The surprise that appears on more bills than any other: NAT Gateway data processing charges. 

If EC2 instances or Lambda functions are accessing S3 through a NAT Gateway instead of an S3 VPC endpoint, every GB of data costs an additional $0.045/GB in NAT Gateway processing fees, on top of any S3 charges. S3 VPC endpoints are free to create and use. The fix takes an afternoon and the savings are permanent.

For finance teams reviewing unexplained S3 spend: the culprit is often not S3 itself but the routing layer around it. CloudZero’s cloud cost monitoring surfaces NAT Gateway-related S3 costs alongside the S3 charges themselves, the only way to see the true cost of your data access patterns.

Serving content directly from S3 to end users? At $0.09/GB for the first 10 TB of internet egress, 1 TB of video or assets costs $90/month in egress charges. Routing the same traffic through CloudFront costs $0 from S3 to CloudFront. CloudFront has its own distribution pricing, but for most content workloads the combined cost is significantly lower than S3 direct egress.

S3 request and retrieval costs

At small scale, request costs are a rounding error. For AI workloads, ML training pipelines reading the same dataset hundreds of times, RAG systems running thousands of vector queries per minute, data pipelines with automated Lifecycle transitions, request costs can exceed storage costs on the same bill.

Three patterns that generate the most unexpected request spend:

How do you estimate your monthly S3 bill?

The AWS Pricing Calculator is the most accurate way to model complex scenarios. For a quick finance-grade estimate: the table below shows pure storage costs at three common scales. In production, the total bill is typically 1.3–2.5x the storage figure once requests and transfer are added.

The finance case study: a company storing 100 TB of archived compliance data in S3 Standard pays $2,305/month, or $27,660/year. The same data in Glacier Deep Archive costs $101.38/month, or $1,217/year. The difference, $26,443/year, is recoverable immediately with a single Lifecycle policy. The AWS Pricing Calculator will model the retrieval costs too, so you can validate the economics before moving data.

Where CloudZero adds value in this analysis: knowing which buckets actually contain cold data, allocated to which team or product. Without that visibility, finance teams are estimating against an S3 bill that says ‘$2,305/month’ with no indication of whether that’s active data, archived data, forgotten AI training runs, or all three.

What are the new S3 storage types for AI workloads?

AWS shipped three storage types in 2024–2025. They’re not replacements for the eight storage classes above, they’re purpose-built additions for AI inference, RAG pipelines, and analytics data lakes. Each creates new billing dimensions that don’t roll up into the standard S3 cost line cleanly, which means teams building AI products are often flying blind on these costs.

S3 Vectors: the AI spend line item most companies can’t see

S3 Vectors launched generally available in December 2025 at re:Invent. It’s the first cloud object store with native support for storing and querying vector embeddings, scaling to 2 billion vectors per index and 10,000 indexes per bucket. AWS claims up to 90% lower cost versus standalone vector databases.

The AI spend angle: if your company is building RAG pipelines, semantic search, AI agent memory, or recommendation systems, your vectors live somewhere, Pinecone, Weaviate, Qdrant, OpenSearch. Each carries a monthly subscription that doesn’t require a board-level conversation to sign but does show up on the cloud bill. S3 Vectors is the AWS-native alternative with materially different economics.

Concrete AI ROI comparison: a workload with 1 million vectors at 1,024 dimensions, roughly 4.7 GB of logical storage, costs approximately $0.28/month in storage on S3 Vectors. 

Pinecone’s starter plan starts at $50/month. For most RAG and agent memory workloads that don’t require sub-10ms query latency, that’s the 90% savings claim translated to a number your finance team can verify.

S3 Vectors integrates with Amazon Bedrock Knowledge Bases for RAG workflows and Amazon SageMaker Unified Studio. Available in 14 AWS Regions. Trade-off: S3 Vectors doesn’t support hybrid (semantic + keyword) search via Bedrock Knowledge Bases, for that, OpenSearch is still required. For pure semantic search and agent memory at scale, S3 Vectors is the cost-sensible default.

The AI spend visibility problem: S3 Vectors costs appear as separate line items under S3 in Cost Explorer, but they blend into the general S3 spend without allocation. If you’re not attributing vector storage costs to the AI products or teams generating them, your AI ROI calculation is missing a real variable. CloudZero’s AI spend intelligence allocates vector index costs to specific models, products, and teams, giving finance the visibility to answer ‘what does our RAG pipeline actually cost per query?’

S3 Tables: the data lakehouse cost layer

S3 Tables has expanded significantly. Purpose-built for tabular data as Apache Iceberg tables, with 3x faster query performance and 10x higher transactions per second compared to general-purpose buckets for Iceberg workloads. If you’re running a data lakehouse for ML feature stores, business analytics, or AI training data management, S3 Tables is the relevant storage layer.

Table-level cost visibility became available in AWS Cost Explorer as of June 2025, making it practical to attribute S3 Tables spend to specific data products or analytics workloads. 

Databricks, Athena, EMR, Spark, and SageMaker Lakehouse all integrate with S3 Tables, one Iceberg layer across your entire analytics stack.

Finance risk to flag: high object churn environments, tables with heavy write traffic, can drive compaction costs above expectations. A table that ingests 30,000 new 5 MB files per month incurs approximately $0.73 in compaction data charges. Model your write patterns before migrating a high-churn workload.

S3 Metadata: actual pricing (and the cost risk most teams miss)

S3 Metadata launched at re:Invent 2024 and expanded in 2025. It automatically captures object metadata into two Iceberg-based tables: a journal table (real-time log of changes) and a live inventory table (complete bucket snapshot refreshed within an hour of changes). Currently available in us-east-1, us-east-2, and us-west-2.

The actual pricing, per the AWS announcement:

The practical value: SQL-queryable real-time metadata instead of batch S3 Inventory or custom Lambda functions. For cost attribution and audit use cases, genuinely useful.

The cost risk: journal records accumulate indefinitely unless you set an expiration policy. For high-churn buckets (frequent uploads and deletions), journal costs can grow significantly. Set the expiration at configuration time, not after you’ve seen the bill.

Management, analytics, and other S3 costs

Several S3 features carry their own charges that accumulate quietly. Most are optional, but some are worth enabling, the data they surface pays for their own cost and then some.

Versioning overhead deserves a line item on every finance review. When versioning is enabled and no Lifecycle rule expires non-current versions, every overwrite creates a new billable copy. 

On buckets with active write traffic, particularly AI training pipelines writing model checkpoints or experiment artifacts, old versions compound fast. A Lifecycle rule expiring non-current versions after 30–90 days is often the highest-return single S3 optimization available. Software engineering tools that write build artifacts to S3 on every run are the most common offender.

13 ways to reduce your S3 bill

Ordered by impact-to-effort ratio. Implementation detail on each in our full S3 cost optimization guide.

How CloudZero helps finance teams understand S3 and AI spend

Your AWS bill says S3 cost you $180,000 last quarter. What it doesn’t say: which team’s AI training pipeline drove $40,000 of that, which product’s vector index grew 200% month-over-month, or whether last quarter’s Glacier retrieval fees were planned or someone’s emergency. Finance teams making decisions at that level of detail are working without the breakdown they need.

This is the visibility gap CloudZero closes. CloudZero’s CostFormation maps S3 spend to the business dimensions that matter, team, product, feature, customer, AI workload, without requiring perfect AWS tagging as a prerequisite. The AnyCost API extends this across multi-cloud environments, not just AWS.

On AI spend specifically: CloudZero treats S3 Vectors, S3 Tables, and S3 Metadata as distinct cost types, not noise in the S3 line item. Finance teams can see vector storage costs by AI product, Iceberg table costs by data team, and training dataset costs by ML experiment, the granularity needed to calculate AI ROI with confidence.

Duolingo, Grammarly, PicPay, Skyscanner and other leading global organizations use CloudZero to answer the questions their AWS bills can’t: how much does our RAG pipeline cost per 1,000 queries? Which team’s versioning overhead is the biggest contributor to S3 spend? What’s our vector storage cost as a percentage of AI revenue? Most teams we work with don’t have a spend-cutting problem. They have a visibility problem. Once they can see S3 costs broken down by team, product, and AI workload, the right optimizations become obvious, without a dedicated cost-engineering sprint.

to see how CloudZero customers allocate S3 and AI storage costs. Not ready yet? Take the self-guided product tour at your own pace, or get started with a free cloud cost assessment to see where your current spend is going.

Frequently Asked Questions about Amazon S3 Pricing

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.