AWS S3 vs Cloudflare R2

S3 is the reference implementation of object storage and the most feature-complete product in this category by a distance. Storage classes from Standard down to Glacier Deep Archive, Intelligent-Tiering, replication, lifecycle rules, event notifications into the whole AWS ecosystem, Object Lambda, S3 Select, Access Points, and IAM permissions granular enough to satisfy any auditor. If you need one of those things, nothing else on this page does it.

R2 does not try to match that. It implements the parts of the S3 API that most applications use, and it charges nothing for egress. That single pricing decision is the entire pitch, and for a large class of workloads it is decisive: serving five terabytes a month out of S3 costs hundreds of dollars, and out of R2 costs zero.

If your bucket is a public read path — images, video, downloads, static assets, a package registry — move to R2. The savings are large, immediate, and the migration is a data copy plus an endpoint change.

If your bucket is deep inside an AWS architecture — feeding Athena, triggering Lambdas, tiering to Glacier, replicating across regions under a compliance policy — stay on S3. The egress you're paying is probably intra-AWS anyway, and the integration is worth more than the storage rate.

The common architecture is both: S3 for the internal data lake, R2 for anything users download. That's not a hedge, it's just matching the tool to the traffic pattern.

AWS charges roughly nine US cents per gigabyte to send data out to the internet, with volume discounts that only matter at large scale. That's about ninety dollars per terabyte. A modest media site serving ten terabytes a month is paying nine hundred dollars just to deliver bytes it already paid to store.

R2 charges nothing. Not a lower rate — nothing. You pay for storage and for operations, and downloads are free regardless of volume or destination.

Our storage hub computes a standard scenario across every provider — 1 TB stored, 5 TB egress, 10 million requests a month, in USD and AUD — and the gap on that shape is not subtle. Egress is where the entire category's economics live, and it's why R2 exists.

Two honest caveats. First, if you put CloudFront in front of S3, your egress is billed at CloudFront rates rather than S3's, which is cheaper — still not free, but the naive comparison overstates the gap for anyone already using a CDN properly. Second, R2's operations pricing is real: Class A operations (writes, lists) cost meaningfully more than Class B (reads). A workload doing enormous numbers of small writes can find that operations, not storage, is its largest line.

For read-heavy public workloads, neither caveat changes the conclusion.

R2 covers the common S3 API surface well, but S3 has fifteen years of features and R2 has not reimplemented all of them. Check this list against your code before committing.

Storage classes are the biggest gap. S3 lets you tier cold data to Glacier at a fraction of Standard pricing, which for a large archive is a bigger saving than R2's free egress. R2's model is simpler with fewer tiers — verify the current options against your retention needs.

Event notifications differ. S3's fan out to Lambda, SQS and EventBridge with a mature configuration model. R2's event story routes into Cloudflare's own ecosystem, which is fine if you're building there and awkward if you aren't.

Also verify: object lock and WORM support if you have a compliance retention requirement; versioning semantics; lifecycle rule expressiveness; conditional request and ETag behaviour on multipart objects; and the granularity of access controls, since R2's token model is simpler than IAM and simpler can mean insufficient.

The practical approach is to run your existing integration test suite against an R2 bucket before migrating anything. Most applications pass unchanged. The ones that don't usually fail on exactly one of the items above, and it's better to find out in CI.

The mechanics are unglamorous and well trodden. Copy the data with rclone or a similar tool, verify checksums, dual-write for a period, cut reads over, then decommission.

Two details make it painless. First, R2 supports migration modes that copy objects on demand the first time they're requested, so you can flip the read path immediately and let the bucket fill in behind you rather than waiting for a bulk copy of many terabytes. Second — and this is the delightful part — AWS does not charge you egress for data leaving to a Bandwidth Alliance partner in the standard way you'd fear, but you should confirm the current terms, because a bulk copy of a large bucket out of S3 can otherwise cost more than a year of R2.

Practical sequence: stand up the R2 bucket, dual-write new objects to both, start the bulk copy of historical objects, verify a sample by checksum, switch reads to R2 behind a feature flag, watch error rates for a week, stop writing to S3, and leave the S3 bucket in place for a month before deleting anything.

Keep your object keys identical. The migration then reduces to changing an endpoint and a set of credentials, and rolling back is changing them again.

S3's real product is not storage, it's the ecosystem. Query it in place with Athena. Train on it with SageMaker. Trigger workflows with EventBridge. Tier it automatically with Intelligent-Tiering. Govern it with IAM policies your security team has already reviewed. If your organisation runs on AWS, S3 is not a component you're choosing — it's the floor.

R2's real product is delivery. It sits next to Workers, which means an object can be read, transformed and served from the edge without an origin request. Combined with Cloudflare Images or a transform Worker, it becomes an image pipeline. Combined with Cache Rules, it becomes an asset host with a global CDN and no bandwidth bill.

For Australian teams there's a latency dimension worth noting. Cloudflare's network reaches Sydney, Melbourne, Brisbane and Perth, so R2-backed content is served locally regardless of where the bucket lives. S3 has an ap-southeast-2 region in Sydney, which gives genuine data residency — a stronger guarantee than edge caching if residency is a compliance requirement rather than a performance preference. We cover that distinction properly on the Australian storage page.

The wider field — Backblaze B2, Wasabi, Tigris, Storj, Hetzner, Scaleway and the rest — is on the storage hub, compared on the same fields.