Why Do AI Agents Cost So Much?
Agents aren't expensive per call — they're expensive per hour. The bill is volume, retries, context, and ungoverned tools, not the token rate.
The real cost drivers
Volume: agents call tools hundreds of times per task.
Retries: loops multiply single failures.
Context: long sessions re-send big contexts.
Ungoverned tools: paid APIs and data called without review.
Why it feels sudden
The per-call cost is tiny; the multiplier is the surprise.
A $12,400 night is 40 retries of one purchase — not one big charge.
The levers
Velocity limits kill the multipliers. Caps bound the totals. Allowlists control the vendors. Model selection and caching cut the rate.
At a glance
| Driver | Lever |
|---|---|
| Volume | Per-agent ceilings |
| Retries | Velocity limits |
| Context | Trimming + caps |
| Tools | Merchant allowlist |
FAQ
Is it the model cost?
Partly — but volume and retries dominate. Rate shopping alone won't fix the bill.
What's the fastest fix?
A velocity limit — it kills the most common multiplier in seconds.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check