www.awsgood.com
BytesDownloaded and PutRequests, and a lifecycle rule that aborts stale multipart uploads after 3 days.People ask "how much does S3 cost per month" expecting one number. There isn't one, because S3 bills five separate meters: storage GB-month by class, request counts by tier, data transfer out, retrieval/early-delete fees, and optional features like Intelligent-Tiering monitoring. Storage is the cheapest of them at SMB scale.
Here's the us-east-1 storage ladder as of the early-2026 rate card. Regional sheets (ap-southeast-2, eu-south-1) run noticeably higher, so pull your own region before quoting anything to a client.
| Storage class | $/GB-month | Min billable duration | Retrieval fee | Sensible for |
|---|---|---|---|---|
| S3 Standard | 0.023 | none | none | live assets, app uploads |
| S3 Intelligent-Tiering | 0.023 / 0.0125 | none | none | unknown access patterns ($0.0025 per 1,000 objects monitored) |
| Standard-IA | 0.0125 | 30 days | $0.01/GB | backups read a few times a quarter |
| One Zone-IA | 0.0100 | 30 days | $0.01/GB | regenerable thumbnails, caches |
| Glacier Instant Retrieval | 0.004 | 90 days | $0.03/GB | archives that must open instantly |
| Glacier Flexible Retrieval | 0.0036 | 90 days | $0.01/GB bulk | compliance dumps |
| Glacier Deep Archive | 0.00099 | 180 days | $0.02/GB | 7-year retention |
Now the same bucket with real traffic on it. A 2 TB media bucket serving a site with 30 million uploads and 4 TB of monthly egress looks like this:
| Line item | Usage | Unit price | Monthly |
|---|---|---|---|
| TimedStorage-ByteHrs (Standard) | 2,048 GB | $0.023/GB-month | $47.10 |
| Requests-Tier1 (PUT/COPY/POST/LIST) | 30 million | $0.005 per 1,000 | $150.00 |
| DataTransfer-Out-Bytes | 4,096 GB | $0.09/GB | $368.64 |
| Total | $565.74 |
Storage is 8% of that bill. Egress is 65%. If you're sizing a budget from the price-per-GB headline, you're modelling the wrong meter — and the gap widens fast once a CDN isn't in front of the bucket.
Email-only budgets get ignored; I've watched a team leave a $40 alarm unread for nine days. AWS Budgets supports budget actions, which attach an IAM deny policy when a threshold trips. That's the layer that saves you.
Create the budget against the S3 service filter first:
# s3-monthly-budget.json
{
"BudgetName": "s3-monthly",
"BudgetLimit": { "Amount": "150", "Unit": "USD" },
"TimeUnit": "MONTHLY",
"BudgetType": "COST",
"CostFilters": { "Service": ["Amazon Simple Storage Service"] },
"CostTypes": { "IncludeCredit": false, "IncludeRefund": false, "UseBlended": false }
}
aws budgets create-budget
--account-id 123456789012
--budget file://s3-monthly-budget.json
Then the action. It needs an execution role that trusts `budgets.amazonaws.com` and can attach the deny policy to your uploader roles:
```bash
# deny-s3-writes.json (the action definition, not the IAM policy)
{
"IamActionDefinition": {
"PolicyArn": "arn:aws:iam::123456789012:policy/S3StopWrites",
"Roles": ["app-uploader", "ci-deploy"]
}
}
aws budgets create-budget-action
--account-id 123456789012
--budget-name "s3-monthly"
--notification-type ACTUAL
--action-type APPLY_IAM_POLICY
--action-threshold 100
--action-threshold-type PERCENTAGE
--approval-model AUTOMATIC
--execution-role-arn arn:aws:iam::123456789012:role/BudgetActionRole
--definition file://deny-s3-writes.json
One caveat that bit me: Budgets re-evaluates roughly three times a day, so this is a 4–8 hour fuse, not a real-time circuit breaker. Pair it with layer 2 for anything that can burn hundreds of dollars in an hour. Also test the deny policy in a sandbox account first — an automatic action that blocks the *deploy* role at 100% of forecast can take production down harder than the bill would have.
## Layer 2: CloudWatch alarms on bytes and requests
S3 publishes `BucketSizeBytes` and `NumberOfObjects` for free, once per day, with roughly a 24-hour lag — good for trend lines, useless for incidents. Per-request metrics have to be switched on, and they carry a CloudWatch custom-metric charge, so enable them on your top 10 buckets by traffic, not all 200.
```bash
aws s3api put-bucket-metrics-configuration
--bucket my-bucket
--id EntireBucket
--metrics-configuration '{"Id":"EntireBucket"}'
# 500 GB downloaded in 6 hours -> page someone
aws cloudwatch put-metric-alarm
--alarm-name "s3-my-bucket-egress-6h"
--namespace AWS/S3
--metric-name BytesDownloaded
--dimensions Name=BucketName,Value=my-bucket Name=FilterId,Value=EntireBucket
--statistic Sum --period 3600 --evaluation-periods 6
--threshold 536870912000
--comparison-operator GreaterThanThreshold
--treat-missing-data notBreaching
--alarm-actions arn:aws:sns:us-east-1:123456789012:ops-pager
Thresholds should be relative to a real baseline, not round numbers on a slide:
| Layer | Signal | Threshold | Fires to |
|---|---|---|---|
| Budgets | S3 cost, monthly | 60% of limit (forecast) | Slack #alerts |
| Budgets | S3 cost, monthly | 100% actual | Pager + IAM deny action |
| CloudWatch | BytesDownloaded, Sum 6h | 500 GB | Pager |
| CloudWatch | PutRequests, Sum 1h | 20× 7-day baseline | Slack #alerts |
| Cost Anomaly Detection | SERVICE = S3 | $25 absolute impact | Email |
Cost Anomaly Detection is the cheapest safety net — it learns seasonality for free and doesn't need you to invent a number. It needs a few weeks of history to be useful, so don't expect it to catch a spike in a brand-new account.
## Layer 3: lifecycle rules, including the one everyone forgets
`AbortIncompleteMultipartUpload` is the single highest-value line in an S3 lifecycle policy. A failed 5 GB upload leaves part files behind that bill at Standard rates and never appear in the console's object list — you only see them via `aws s3api list-multipart-uploads`.
```json
{
"Rules": [
{
"ID": "tier-and-clean",
"Filter": { "Prefix": "" },
"Status": "Enabled",
"Transitions": [{ "Days": 0, "StorageClass": "INTELLIGENT_TIERING" }],
"AbortIncompleteMultipartUpload": { "DaysAfterInitiation": 3 },
"NoncurrentVersionExpiration": { "NoncurrentDays": 30 },
"Expiration": { "ExpiredObjectDeleteMarker": true }
}
]
}
```bash
aws s3api put-bucket-lifecycle-configuration
--bucket my-bucket
--lifecycle-configuration file://lifecycle.json
Objects under 128 KB don't get transitioned by lifecycle rules at all — they stay in the source class and keep billing there. If your bucket holds millions of small JSON files, moving them to Glacier buys you nothing but a longer restore time.
## Field note: the 03:40 alert that wasn't an alert
Our CI artifact bucket went from 90 GB to 1.3 TB in eleven days. The daily spend graph went from about $2.10 to $31, and the only reason we looked was a client reviewer asking why the account total doubled. Versioning was on — correct, we wanted it — but nothing expired noncurrent versions, so every one of ~3,000 builds kept its full 400 MB artifact forever.
The first config was wrong in a way S3 happily accepts. I'd written the rule with `"Filter": {"Prefix": "builds/latest/"}` instead of `"Prefix": "builds/"`. `put-bucket-lifecycle-configuration` returned HTTP 200 and the rule matched zero objects. S3 will not warn you about a rule that matches nothing.
Diagnosis was three commands. `aws s3api list-object-versions --bucket ci-artifacts --prefix builds/ --query 'length(Versions)'` returned 6,412 versions where I expected ~40. A CloudWatch `get-metric-data` call on `BucketSizeBytes` showed a clean staircase starting the day versioning was enabled. And when I tried to read request metrics I'd assumed were already on, I got this back:
An error