API payload size and cost: what every extra kilobyte is worth

Most teams size their API by latency and by CPU. Very few size it by bytes on the wire, which is odd, because bytes on the wire are the part the cloud bills directly. A response body is charged as egress the moment it crosses from your region to the internet, and unlike compute there is no way to amortize it with a reservation or a savings plan. The meter is linear and unforgiving.

The unit economics of a kilobyte

Start with the base number. AWS charges 0.09 dollars per GB for the first 10 TB of internet egress in US regions. One GB is 1,048,576 KB, so a single kilobyte costs about 0.0000000858 dollars. Multiply by a billion and you get roughly 86 dollars. That is the number worth memorizing: each kilobyte of response body costs about 86 dollars per billion calls per month.

At a million calls a month nothing here matters. At a billion, the difference between a lean 2 KB response and a lazy 20 KB one is about 1,550 dollars every month, which is a full-time engineer's cloud budget spent on whitespace and unused fields.

Where the bloat comes from

The recurring offenders are predictable. Serializing an entire database row when the client needs four fields. Returning nested objects the caller discards. Pretty-printed JSON with two-space indentation, which can add 10 to 20 percent to a deeply nested document for zero benefit in a machine-to-machine call. ISO timestamps with microsecond precision and timezone offsets where a Unix integer would do. Null fields serialized explicitly. Base64-encoded binary inline, which inflates by 33 percent before any other consideration. And the biggest one: unbounded list endpoints that return every record because nobody implemented pagination.

Compression changes the arithmetic

JSON compresses extremely well because it is repetitive by nature. Gzip typically reduces a JSON body by 70 to 85 percent, and Brotli at a moderate quality level gets another 10 to 15 percent beyond gzip on text. A 20 KB response that compresses to 3 KB drops the billion-call cost from about 1,717 dollars to about 258 dollars. The CPU cost of compressing is real but small: gzip at level 6 on a 20 KB body takes well under a millisecond on a modern core, which on a 0.0432 dollars per hour m7g.large works out to a rounding error against the transfer saved. The trade-off is analyzed in compression to reduce cloud cost.

Headers count too

A response body is not the whole payload. HTTP headers travel on every response, and verbose CORS headers, long cache-control directives, tracing headers, and oversized cookies can add 500 bytes to 2 KB per response. For a chatty API making many small calls, headers can exceed the body. A 200-byte JSON response with 1.5 KB of headers is spending 88 percent of its bytes on metadata. HTTP/2 header compression helps between the client and the edge, but if your origin is behind a load balancer that re-expands headers, the internal hop pays full price.

The hidden multipliers

Payload cost rarely stops at one hop. A response that travels from a service in one availability zone to an API gateway in another pays 0.01 dollars per GB each way before it ever reaches the internet. If a request fans out to five internal services and each returns a full object that the aggregator trims, you pay cross-zone charges on the untrimmed version five times. This is the pattern described in chatty microservices network cost. Trimming at the source, not at the edge, compounds the saving across every internal hop.

A practical payload budget

Give each endpoint a byte budget the way you would give it a latency budget. A reasonable starting point: list endpoints capped at 100 items with a hard page size, detail endpoints under 4 KB uncompressed, compression mandatory above 1 KB, no pretty-printing outside development, null fields omitted, and binary served as a signed URL to object storage rather than inlined. Measure the p50 and p99 response size in your metrics alongside latency. Once the number is visible in a dashboard, it stops growing quietly.

Price it before it ships

The reason payload bloat survives so long is that nobody sees the bill attributed to it. The egress line says only that you moved 4.66 TB, not that 60 percent of it was unused fields. Estimating traffic volume for a new endpoint during design, against real per-GB rates, makes the trade-off concrete while the schema is still easy to change. Price your API infrastructure against the resource catalog before the schema is frozen.

FAQ

What to do next

Know what your traffic costs before it ships. C3X prices your Terraform against a live resource catalog. Start with the quickstart.