Amazon S3

What Amazon S3 actually is

Amazon S3 launched on Pi Day, March 14, 2006 — the first generally available AWS service, announced in a single short paragraph as "storage for the Internet." Twenty years later it stores more than 500 trillion objects across hundreds of exabytes and serves over 200 million requests per second globally. It is object storage: you PUT a blob with a key, you GET it back, and everything else — versioning, lifecycle rules, replication, access policies, event notifications, querying — hangs off those two primitives. The API you wrote against in 2006 still works unchanged, and that backward compatibility is a large part of why "S3-compatible" is an industry standard rather than a vendor feature.

The storage class ladder is the core of the product. S3 Standard for frequently accessed data; Intelligent-Tiering that moves objects between access tiers automatically; Standard-IA and One Zone-IA for infrequent access; Express One Zone for latency-sensitive single-AZ workloads; and the Glacier family (Instant Retrieval, Flexible Retrieval, Deep Archive) for archives. All Regional classes replicate across a minimum of three Availability Zones by default and are designed for 99.999999999% (11 nines) data durability — S3's continuous auditor microservices inspect every byte across the fleet and trigger repair on any sign of degradation.

One design choice deserves emphasis: objects are immutable writes with a key, and every advanced behavior — versioning, replication, retention — is layered on by the platform rather than by your code. That is why S3 scales the way it does and why the API has never broken: there is very little to get wrong at the storage layer, and AWS has spent twenty years making the hard parts (consistency, durability auditing, tier transitions) invisible.

Around that core sit two decades of features: strong read-after-write consistency, versioning (since 2010), Object Lock for WORM compliance, S3 Replication, Inventory, Access Points, Access Grants, Transfer Acceleration, server-side encryption with KMS or customer keys, static website hosting, event notifications into Lambda/SQS, and newer shapes like S3 Tables and S3 Vectors. VPC endpoints and S3 Access Points handle the network and multi-tenant access patterns that used to require custom engineering.

Pricing: what you really pay

Pay only for what you use, with no minimum charge. In US East (N. Virginia), from the published price list: S3 Standard storage is $0.023 per GB-month for the first 50 TB, $0.022 for the next 450 TB, and $0.021 above 500 TB. Requests: $0.005 per 1,000 PUT/COPY/POST/LIST calls, $0.0004 per 1,000 GET requests. Data transfer out to the internet runs $0.09 per GB for the first 10 TB after the 100 GB/month free allowance aggregated across AWS services.

The cheaper classes change the arithmetic substantially: Standard-IA at $0.0125 per GB-month (plus per-GB retrieval fees), Glacier Instant Retrieval at $0.004, Glacier Flexible Retrieval at $0.0036, and Glacier Deep Archive at $0.00099 per GB-month — a 23x spread from top to bottom of the ladder. Region matters: the same Standard GB costs $0.0245 in Frankfurt, $0.025 in Singapore and $0.0405 in São Paulo. A worked example from the price list: 1 TB in S3 Standard with 1M GETs and 100k PUTs runs about $24.45/month in us-east-1.

The classic free tier framing has changed for new accounts: since July 15, 2025, new AWS customers receive up to $200 in Free Tier credits applicable to S3 and other eligible services, with a free plan available for 6 months and credits usable within 12 months. What consistently surprises people is not storage cost but the meters around it: PUT/GET request counts at high frequency, data retrieval fees on IA and Glacier classes, replication traffic, Intelligent-Tiering monitoring charges, and above all egress. Backups that are cheap to store and expensive to restore are a trap S3 shares with every cloud archive; price the restore path before you archive.

We have not benchmarked S3 against Backblaze B2 or Wasabi in our /benchmarks runs, so no throughput claims here. What we can say is that the pricing model is legible once you know the four billable dimensions — storage, requests, retrieval and transfer — and that the per-GB storage rate has dropped roughly 85% since 2006.

Performance & infrastructure

S3 runs across 123+ Availability Zones in 39 AWS regions (AWS's own birthday post cites 123 AZs; the current infrastructure page lists 124), with hundreds of edge locations for Transfer Acceleration and CloudFront-fronted delivery. Performance is designed around high aggregate throughput rather than single-stream latency: multipart uploads, request parallelism and S3 Transfer Acceleration are the levers, and the service scales to absorb traffic spikes without provisioning. Express One Zone exists precisely for the workloads where single-digit-millisecond access inside one AZ beats Regional replication.

Reliability is the product's strongest dimension and its most honest marketing. Eleven-nines durability is a design target backed by the cross-AZ redundancy and the continuous byte-level auditing; strong read-after-write consistency (delivered in 2020) closed the last big correctness gap. The SLA is more modest than the durability story: 99.9% monthly uptime for S3 Standard, Express One Zone and the Glacier Flexible/Deep Archive classes, and 99.0% for Intelligent-Tiering, Standard-IA, One Zone-IA and Glacier Instant Retrieval, with service credits of 10%, 25% or 100%. Durability protects against data loss; the SLA covers availability — they are different promises and AWS is clear about which is which.

The track record is not spotless, and honesty requires saying so: the February 2017 us-east-1 outage took large parts of the internet down for hours, and S3's default-permissive history (public buckets by default in the early years) contributed to a decade of data leaks. Both are addressed by current defaults — Block Public Access is on by default now — but they belong in any twenty-year account.

Where it falls short

Egress is the standing complaint and it is structural: getting data out of AWS costs real money per GB, and at scale it reshapes architecture (people cache with CloudFront, replicate selectively, or accept the bill). Restoring a Glacier archive also takes minutes-to-hours and costs per-GB retrieval — fine for disaster recovery, wrong for anything interactive.

Request pricing punishes chatty designs. A workload doing billions of small GETs can spend more on requests than on storage; the fix is batching, caching, or a different class, but the bill does not warn you first. Intelligent-Tiering's monitoring charge and the per-object metadata overhead on archived objects (8 KB plus 32 KB billed at archive rates) are the kind of details that only surface in deep bills.

Configuration complexity is real despite the simple API. Bucket policies, IAM policies, ACLs, Block Public Access, Object Ownership, VPC endpoints and KMS key policies interact; misconfiguration remains the dominant S3 security incident pattern. Encryption is on by default today, but key management choices still matter. And there is no free compute attached — S3 is storage, and anything dynamic runs elsewhere.

Who should pick this

Our pick when durability, API permanence and ecosystem matter more than the last fraction of a cent per GB. It is the default for AWS-native architectures, static asset delivery behind CloudFront, data lakes (with S3 as the storage layer for Athena, Glue, EMR and every ML framework that reads from object storage), backup targets, and log archives with lifecycle policies to Glacier. If you are building anything that other tools will integrate with, the S3 API is the safest bet in storage.

We would also pick Deep Archive deliberately for compliance retention: $0.00099 per GB-month with Object Lock is about as cheap as audited retention gets.

We would not pick S3 for pure price-per-GB at moderate scale — Backblaze B2 and Wasabi are cheaper for straightforward storage and backup — nor for teams trying to stay cloud-portable on price. The correct strategic view in our /category/cloud-object-storage category: S3's premium buys you the API standard, the durability engineering and twenty years of backward compatibility.

Key takeaways

• Launched March 14, 2006 — the first GA AWS service; the 2006 API still works unchanged, which is why S3 compatibility is an industry standard.
• Designed for 11 nines (99.999999999%) durability with cross-AZ replication by default; strong read-after-write consistency.
• S3 Standard in us-east-1: $0.023/GB-month (first 50 TB), $0.005 per 1,000 writes, $0.0004 per 1,000 reads; egress $0.09/GB after a 100 GB monthly free allowance.
• The class ladder spans 23x: Standard-IA $0.0125/GB-month down to Glacier Deep Archive $0.00099/GB-month.
• SLA is 99.9% for S3 Standard and 99.0% for IA/tiering classes — durability and availability are different promises.
• Watch the meters around storage: request counts, IA/Glacier retrievals and egress dominate surprise bills; price the restore path first.

Sources

• https://aws.amazon.com/s3/pricing/
• https://aws.amazon.com/s3/sla/
• https://aws.amazon.com/s3/faqs/
• https://aws.amazon.com/about-aws/global-infrastructure/
• https://aws.amazon.com/shield/pricing/
• https://aws.amazon.com/jp/blogs/aws/twenty-years-of-amazon-s3-and-building-whats-next
• https://cloud-credits.org/prices/s3/us-east-1
• https://upstash.com/blog/aws-s3-pricing-explained-2026
• https://howmakersbuild.com/tools/aws-s3/pricing
• https://en.wikipedia.org/wiki/Amazon_Web_Services