Why a million small objects cost more than the same bytes in one file

A size floor and metadata overhead are different meters

A minimum billable object size replaces the charged size when an object is below a class threshold. S3 Standard-IA, One Zone-IA, and Glacier Instant Retrieval bill objects smaller than 128 KB as 128 KB. S3 Glacier Flexible Retrieval and Deep Archive instead add 40 KB of metadata to every archived object. [1]

AWS splits that 40 KB into 8 KB billed at S3 Standard rates and 32 KB billed at the archive-class rate. One million archived objects therefore add 40 GB of metadata in the decimal KB-to-GB arithmetic used by the AWS records, even when their logical payload already totals 1,000 GB. [1] [2]

The dataset also contains GiB and KiB offers. This explainer keeps each published unit system visible and does not silently treat KB as KiB or GB as GiB. [3] [4]

The same 1,000 GB can carry $0.99 or $216.67 of Deep Archive storage

The worked workload holds logical payload at 1,000 GB for one full month in S3 Glacier Deep Archive in us-east-1. Average object size divides the payload evenly, the decimal dataset units are kept as published, and replication, versioning, requests, retrieval, temporary copies, and egress are excluded from this storage-only table. [2] [1]

The payload line is always 1,000 GB x $0.00099 = $0.99. Object size changes only the object count and the two metadata lines. At a 1 MB average there are one million objects, so the backup carries 40 GB of pure metadata and costs $1.21 after rounding. [2] [1]

Below 9.80 KB, Deep Archive is more expensive than S3 Standard

For this 1,000 GB us-east-1 comparison, S3 Standard costs 1,000 x $0.023 = $23.00 in the first storage band and has no minimum billable object size. Deep Archive costs $0.99 for payload plus $215.68 divided by average object size in KB for its split metadata. [2] [1]

Setting $0.99 + $215.68 / size equal to $23.00 gives a crossover at 9.799 KB, rounded to 9.80 KB. Below that average, the nominally cheapest class is the more expensive of these two storage tiers. This is a storage-only crossover, not a restore winner, because retrieval, requests, temporary-copy storage, and egress remain separate. [2] [1]

A read-only restore exposes the request cost of small objects

This restore keeps the archived source unchanged and restores the same 1,000 GB to the public internet using Deep Archive Standard retrieval. It assumes one RestoreObject request and one GET per object, the full 100 GB AWS internet-transfer allowance remains, and the temporary S3 Standard copy is kept for 7 days in an illustrative 30-day month. [2] [1] [5]

The byte-based subtotal is constant: $20.00 retrieval, $81.00 egress, and $5.37 temporary-copy storage, or $106.37 after rounding. Object count changes the request subtotal at $0.10 per 1,000 Standard restore requests plus $0.0004 per 1,000 GET requests. Real tools can issue additional HEAD, LIST, range, and retry operations, so the one-restore-plus-one-GET assumption is a lower-shape model rather than a universal request count. [2] [1]

Early deletion does not apply and contributes $0.00. A read-only restore can charge retrieval, restore and GET requests, temporary-copy storage, and egress, but it does not delete, overwrite, move, or transition the archived source. [1] [5]

Egress and requests decide the restore, not the storage-rate headline

A storage-only crossover answers which tier holds the bytes more cheaply. It does not answer which provider restores them more cheaply. Backblaze B2, Cloudflare R2, and Wasabi expose three different restore boundaries: a proportional egress allowance with a published overage, no internet-egress charge with paid operations, and a fair-use policy with no published excess rate. [6] [8] [9]

The examples below use 1,000 GB of monthly average or active storage as the provider rule requires. They keep egress and request lines separate, and they do not add early-deletion charges because every restore remains read-only. [6] [8] [9] [5]

All 18 vendors keep their published rule or honest state

The 61-offer provider dataset covers 18 vendors. A positive floor appears only where the captured record publishes one; not published remains unknown rather than zero, and does not apply is reserved for an explicit absence. Class-specific rules stay class-specific. [1] [10] [11]

Amazon S3, DigitalOcean Cold, Hetzner, Storj, and Telnyx have a positive billable-size rule in the current dataset. Amazon also has the separate 40 KB archive metadata meter. Backblaze, Akamai, and Filebase explicitly publish no positive floor. The remaining ten vendors stay not published, including the unresolved Wasabi conflict documented below. [1] [3] [7] [20] [25]

Conflicts and look-alike rules stay visible

The provider dataset is the rate and state input for this explainer. The DigitalOcean Cold mismatch was reconciled on 2026-07-29, so the canonical record now carries the documented 128 KiB storage and read floors. The separate Wasabi claim remains unresolved and is not silently promoted into arithmetic. [9] [12] [3]

Upload eligibility, tiering eligibility, and recommended object size are also rejected as substitutes for a billable-size rule. They can affect whether an upload or tiering action is supported, but they do not prove how stored bytes are billed. [14] [1] [17]