build.allanninal.dev

Part 6 of 7 · Content repurposer series ~3 min read

What the content repurposer costs

This system only does work when you hand it a long piece. There’s no clock ticking in the background, no always-on anything. When a piece arrives, it reads it, pulls the points, drafts the posts, and goes quiet again. The cost is almost all model calls — the reading and the writing — and even those are cheap because the model used is the small, fast one. At typical SMB volume, the bill is a few dollars a month, fixed cost essentially zero.

Key takeaways

  • Around $3/month at typical SMB volume (about eight long pieces a month).
  • Fixed AWS cost is essentially zero. No always-on compute, no NAT Gateway, no API Gateway.
  • The cost is dominated by Bedrock — pulling points and drafting each clip.
  • Embeddings and S3 Vectors are a small slice; everything else is pennies.
  • At 20 pieces a month the bill is around $6. At 40 it’s around $12.

Cost at three volumes

Monthly cost at three piece volumes, broken out by component A vertical stacked-bar chart showing monthly cost in US dollars at three volumes of long pieces repurposed per month. The leftmost bar represents eight pieces a month and shows a total around $3, dominated by the Bedrock slice (pulling points and drafting each clip) with a small slice for embeddings and S3 Vectors, a tiny fixed slice, and a small everything-else slice. The middle bar represents twenty pieces a month and shows a total around $6, with the same shape — Bedrock grows roughly linearly with the number of pieces because each piece means more points to score and more drafts to write. The rightmost bar represents forty pieces a month and shows a total around $12, with Bedrock still dominant; embeddings and the everything-else bucket grow with it but stay smaller in absolute terms. Below the chart is a legend explaining the four sections of each bar: Bedrock (pulling points and drafting), embeddings plus S3 Vectors (one fingerprint per passage), AWS Budgets and Secrets Manager (small fixed amounts), and an everything-else bucket for Lambda runtime, DynamoDB on-demand, S3, EventBridge Scheduler, SES, and CloudWatch. A note at the bottom: the model calls are the dominant cost — and even those are a fraction of a cent per draft. $0 $5 $10 $15 $20 8 pieces ~$3 20 pieces ~$6 40 pieces ~$12 Bedrock (pulling points + drafting clips) Embeddings + S3 Vectors (one per passage) AWS Budgets + Secrets Manager (fixed) Everything else (Lambda, DDB, S3, Scheduler, SES, CloudWatch) The model calls are the dominant cost — and even those are a fraction of a cent per draft.
Fig 6. Monthly cost at three piece volumes. Bedrock is the dominant slice because reading and drafting are the real work; embeddings and the everything-else bucket are small. The cost scales with how many pieces you repurpose, not with a clock running in the background.

Where the dollars actually go

Bedrock (the bulk). Two jobs cost real money, and both are model calls. First, pulling points: scoring each passage in a piece for how postable it is. A 1,800-word post is maybe thirty short passages, so thirty cheap reads. Second, drafting: writing each chosen point into a thread, short posts, and a caption, then the source-check rewrite if a draft drifted. That’s the larger share. Both run on Haiku 4.5, the small fast model, so even a piece that fans out into a dozen drafts costs a few cents. The harder pieces that go to Sonnet 4.6 cost more per call, but they’re the minority. At eight pieces a month, Bedrock is a couple of dollars.

Embeddings + S3 Vectors. Each passage gets one fingerprint, made once with Titan Text Embeddings V2 and stored in S3 Vectors. A piece of thirty passages is thirty small embedding calls and thirty tiny vectors. Storing and searching them is cheap. A small slice of the bill, growing slowly with the number of pieces.

Lambda runtime. The intake, the points lane, the drafter, the Function URL handler for the buttons, the scheduler jobs. None run long. The drafter is the heaviest because it waits on model calls, but waiting on Bedrock is billed as Bedrock time, not Lambda time. Lambda lands under a dollar at all three volumes.

DynamoDB on-demand. A couple of small tables: the draft state and the cr-audit log. Writes are approvals, edits, and skips; reads are when you open the desk. Pennies a month at any of these volumes.

S3 + storage. The cleaned source pieces, the raw inbound from forwarded transcripts, the drafts. A few MB total at SMB volume. Effectively free.

SES + Scheduler. SES inbound for the transcript lane and outbound for any email summaries: a few cents a year at this scale. EventBridge Scheduler for the weekly drip and the one-off jobs: pennies.

What doesn’t cost money

  • API Gateway. Replaced by Lambda Function URLs for the paste-a-link form and the approve/edit/skip buttons.
  • NAT Gateway. Nothing is in a VPC. No NAT, no $32/month minimum.
  • Always-on compute. No EC2, no Fargate. The system does nothing between pieces.
  • A big always-on index. S3 Vectors is pay-for-what-you-store, not a running cluster. No vector database humming when no one’s repurposing anything.
  • Heavy-model-by-default. Haiku 4.5 does the routine reading and drafting. Sonnet 4.6 fires only on the harder pieces, so you don’t pay for the big model on every clip.

How the cost scales

The bill tracks how many long pieces you feed it, because every piece means passages to score and drafts to write. Bedrock and embeddings grow roughly linearly with piece count; Lambda and DynamoDB grow with them but stay small. There’s no background cost that grows on its own — a month where you repurpose nothing costs essentially nothing. So the bill at 80 pieces a month is around $24, and at 160 it’s around $48. Past those volumes you’re running a content operation, not a side system, and you’d batch the model calls and cache embeddings harder — but those are tunings, not redesigns.

Set an AWS Budgets alarm at $15/month so an unusual spike (a giant transcript, a runaway retry loop) pages you before the bill matters. The normal-volume bill stays well under that ceiling.

Last post in the series: the engineering reference. Same system, drawn for engineers — service names, Lambda inventory, IAM scopes, the S3 Vectors index, DynamoDB schemas, and EventBridge Scheduler config.

All posts