The headline says the world will spend $206 billion on AI agents this year. Your bill is not that. Your bill is nine line items, most of them small, one of them large, and it almost always lands between roughly $80 and $1,000 a month — governed almost entirely by how many tokens your product moves and whether you turned on caching.
Here's the whole budget, with an honest range and the cheapest cut on each line. Read the table, then fix the lines in order of size — because the single most common founder mistake is trimming a $12 line while a mis-configured $200 line runs unattended.
The one number that dominates: inference#
Model tokens are usually more than half the bill, so this is where the first cut has to land. Two levers do almost all the work, and both are configuration, not code:
- Drop the default to a workhorse tier for the calls that don't need frontier reasoning. Just measure cost per completed task, not price per token first — a cheaper model that needs a retry isn't cheaper.
- Turn on prompt caching. On an agent loop with a stable system prompt and tool set, you're otherwise re-billed for the same thousands of tokens on every step. It's the same lever that decides self-host vs API on open weights, and it stacks with — but isn't the same as — context editing.
Fix these two before you touch anything else on this list. Nothing else moves the total as much.
The second-biggest line: runtime#
A sandbox you leave running 24/7 for a bursty agent is money on fire. Start on a managed, pay-per-use runtime and sleep idle instances; self-hosting only wins past a real volume break-even. The Cloud Run vs E2B vs Modal vs Fly comparison has the per-second math.
The small lines: leave them alone until they grow#
These are where founders feel thrifty and are wrong to spend the attention:
- Observability — free tier or self-hosted OSS is plenty at solo scale; here's the cheapest way to see your token bill.
- Vector / memory — run pgvector on the Postgres you already have before buying a managed vector DB, and remember that memory has its own read/write token cost.
- Email — the free tier covers you until sends genuinely dominate.
- Payments — meter machine spend per-call and keep card rails for human purchases.
- Evals — run them inside the CI you already pay for; the marginal cost is zero.
The rule#
Cut in order of size. Inference first, runtime second, everything else only when a real number — not anxiety — forces it. If you want the component-by-component choices behind this budget, the twelve-decision stack guide is the map; this is what it costs to run.
