How Is AWS S3 Priced? What Storage and Data Transfer Cost per Month

The short version
- S3 money comes in four parts: storage fees + request fees + data transfer fees + add-on feature fees (retrieval, monitoring, replication, batch operations). Everyone calculates the first two; the money usually leaks out on the last two.
- us-east-1 Standard storage is $0.023/GB-month, internet egress is $0.09/GB (first 100GB each month free). Store 1TB for a month, have 500GB downloaded, and it's about $60 a month — roughly 400+ RMB. Drop that same 1TB down to Deep Archive via lifecycle rules and storage costs less than $1.10. Unit prices vary by region and storage class; refer to the official AWS pricing page.
- Cost reduction ranked by ROI: lifecycle rules > purge noncurrent versions and incomplete multipart uploads > put CloudFront in front > turn off unnecessary cross-region replication > merge small files. The first two take half an hour to configure, and most buckets can have 40%+ of their storage cost cut.
Cost-cutting levers ranked by ROI: which cut to make first
| Rank | Lever | Typical savings | Implementation cost | Main side effects |
|---|---|---|---|---|
| 1 | Lifecycle transition to IA / Glacier | Storage cost down 45%–95% | Configure a rule once, half an hour | Retrieval fees, minimum storage duration |
| 2 | Purge noncurrent versions + clear incomplete multipart uploads | 10%–40%, depending on the case | 10 minutes to investigate | Deleting the wrong version is irreversible |
| 3 | Front with CloudFront, origin fetch over the internal network | Egress cost down 60%+ | Half a day; requires changing domain and certificates | Useless if cache hit rate is low |
| 4 | Turn off unnecessary cross-region replication | Drops linearly with replication volume | 10 minutes | Disaster recovery capability goes to zero |
| 5 | Merge small files, reduce PUT frequency | Request cost down an order of magnitude | Requires changing business code | Largest amount of change |
This ranking isn't by absolute dollars saved, but by "configure it once and forget it, no code changes needed." Lifecycle rules rank first because they take effect immediately on existing data and continuously on new data, with no rollback cost. CloudFront saves more, but you have to change DNS, request certificates, and handle origin authentication — that's half a day of work all told, and it's easy to trip up on cache keys.
How storage and data transfer costs are actually calculated
Storage unit prices aren't complicated; the traps are in the billing rules. First, the comparison table (us-east-1, 2026 prices, non-China regions):
| Storage class | Unit price $/GB-month | Minimum storage duration | Minimum billable object | Good for |
|---|---|---|---|---|
| S3 Standard | 0.023 | None | None | Hot data, website static assets |
| Standard-IA | 0.0125 | 30 days | 128 KB | Read once every month or two |
| One Zone-IA | 0.01 | 30 days | 128 KB | Rebuildable thumbnails, transcoding intermediates |
| Glacier Instant Retrieval | 0.004 | 90 days | 128 KB | Archives that need millisecond retrieval |
| Glacier Flexible Retrieval | 0.0036 | 90 days | 40 KB | Backups, minute-level retrieval |
| Glacier Deep Archive | 0.00099 | 180 days | 40 KB | Compliance retention, 12-hour retrieval |
The unit price is only the starting point; billing rounds each object's actual footprint up. Put a 1KB icon in IA and you're still billed for 128KB — storage cost inflated 128x. We've seen a bucket stuffed with over 900,000 2KB placeholder images where the bill didn't drop after moving to IA — that's why.
- 代购 SSL 证书 · DigiCert / GeoTrust / RapidSSL 单域名与通配符
- Wildcard 通配符 SSL 证书安装部署续费申请
- HTTPS证书申请 · SSL证书安装部署 · 证书续费
Then there's data transfer: three paths, three prices — don't mix them together:
- S3 to EC2 in the same region: free, over the internal network.
- S3 to CloudFront origin fetch: free. This is the core of why the CloudFront approach saves money.
- Cross-region replication: $0.02/GB, and you pay for storage on both ends.
- Internet egress: $0.09/GB (first 10TB), then tiered down.
Before you calculate, find out exactly how much is in the bucket — don't trust the rough number on the console home page:
aws s3 ls s3://edu-media-prod --recursive --summarize --human-readable | tail -n 3
The output looks like this:
Total Objects: 1284930
Total Size: 3.9 TiB
One more note for teams in China: pulling files from the mainland directly to us-east-1 often drops below 200KB/s in practice — large files are basically unusable. Worse, CloudFront **has no edge locations in mainland China**, so don't expect a CDN layer to fix mainland access speeds; what it fixes is origin fetch and transfer costs for overseas users.
## Real-world testing and pitfalls: version buildup plus incomplete multipart uploads tripled the bill
In November 2025, a client running online courses came to us saying their S3 bill had climbed from just over $200 a month to over $700. I first looked at the Cost Explorer breakdown: the storage portion had gone from $28 to $91, and the rest was transfer.
Logging into the console and opening the Management tab, the object count hadn't grown much, but the storage volume didn't add up. My first thought was versioning — sure enough, it was Enabled, and the bucket had been around for almost two years with nobody cleaning up noncurrent versions.
The first attempt at configuring a lifecycle rule threw an error:
An error occurred (MalformedXML) when calling the PutBucketLifecycleConfiguration operation:
The XML you provided was not well-formed or did not validate against our published schema
At the time the JSON had both `"Prefix": "courses/"` and `"Filter": {"Prefix": "courses/"}`. In the new schema, `Prefix` is a legacy field, and having it alongside `Filter` is an outright validation failure. The equivalent trap in the Chinese-language console: in the filter section of the Lifecycle configuration page, prefix and tag are an either/or dropdown — pick the wrong one and there's no error, but the rule only applies to one tag, so it looks configured but never matches.
After fixing it, I added `NoncurrentVersionExpiration`, then checked for incomplete multipart uploads:
```bash
aws s3api list-multipart-uploads --bucket edu-media-prod --query 'length(Uploads)'
It returned `47`. The earliest `Initiated` was 2025-08-14; all of them were remnants of interrupted course video uploads, taking up 210GB and billed at the Standard tier the whole time. Interrupted uploads are common on flaky networks, and the SDK won't clean them up for you.
The corrected, complete rule looks roughly like this:
```json
{
"Rules": [
{
"ID": "archive-and-clean",
"Filter": { "Prefix": "courses/" },
"Status": "Enabled",
"Transitions": [
{ "Days": 30, "StorageClass": "STANDARD_IA" },
{ "Days": 120, "StorageClass": "GLACIER_IR" }
],
"NoncurrentVersionExpiration": { "NoncurrentDays": 30 },
"AbortIncompleteMultipartUpload": { "DaysAfterInitiation": 7 }
}
]
}
```bash
aws s3api put-bucket-lifecycle-configuration
--bucket edu-media-prod
--lifecycle-configuration file://lifecycle.json
After running it and waiting for the next day's S3 Storage Lens data, storage dropped from 3.9TB to 1.6TB, and a month later it settled at just over $30 in storage costs. What actually saved money in this whole process came down to two things: clearing 2.1TB of noncurrent versions and clearing 210GB of remnants — neither had anything to do with lifecycle transitions.
## When not to use this approach
**For cold data with second-level retrieval requirements, don't stuff it into Glacier Flexible Retrieval or Deep Archive.** Retrieval takes minutes to 12 hours and incurs retrieval fees; one production incident would be enough to wipe out a whole year's savings.
**For buckets with a huge number of very small objects (say hundreds of thousands of 20KB thumbnails), be cautious about enabling Intelligent-Tiering.** Intelligent-Tiering charges a monitoring fee per object count, $0.0025 per 1,000 objects/month — 1.28 million objects is over $3 a month, which may be more than what you save on small objects.
**Don't just turn off versioning to save money.** Once it's off, all noncurrent versions become unrecoverable — an accidental delete is a real delete. The right approach is to keep versioning and use `NoncurrentVersionExpiration` to control how many days noncurrent versions live.
If I were actually doing this, I'd go in this order:
1. Pull the real storage volume and cost breakdown with `aws s3 ls --summarize` and Cost Explorer — don't guess.
2. Configure `AbortIncompleteMultipartUpload` first — zero risk, immediate results.
3. Then configure `NoncurrentVersionExpiration`, watch for a week to confirm nothing depends on rolling back.
4. Finally, do storage class transitions by prefix according to access patterns — one prefix at a time.
If nobody on the team is watching this area, or the bucket
Leave a Reply