Cost per call, metered
Every request carries a measured cost, attributed to a customer and an intent, visible on the same dashboard as latency.
← The K-Framework
★ Amplification · 10× ThroughputLayer 10 of 16Visual guide
PILLAR B · LAYER 04 · B.04Token Economics.
Optimize cost, context, compute.
What a CEO/CTO needs to know
If no one can tell you the cost per request before the monthly bill arrives, the unit economics are running blind, and they may already be upside down.
Cost per intent metered against a budget, with an alert that fires before spend crosses the line.
[WHAT IT IS]
The engineer’s view, in plain language.
The unit economics of an LLM feature are not "infinite tokens, never look at the bill." We model cost per intent, alert on regressions, and rotate to a smaller model when it saves money with no quality drop. Cost discipline is part of the architecture.
[HOW WE BUILD IT]
What “done right” looks like.
Budgets and regression alerts
A per-customer cost budget and an alert when a PR pushes cost up, so an 18k-token prompt change is caught in review, not in the invoice.
Quarterly model review
Is there a newer cheaper model, a distilled local fallback, or RAG that beats prompt-stuffing? Reviewed on a cadence, not after the crisis.
[MATURITY LADDER]
Where does your build sit?
Four rungs from absent to production-grade. Level 3 is the target, and the only one that survives a real production incident.
L0
Absent
Cost is unknown until the monthly bill. No attribution, no budget.
L1
Ad-hoc
Total spend is watched, but not per intent or per customer.
L2
Managed
Cost is metered, but there is no budget gate or regression alert.
L3Target
Production-grade
Cost-per-intent budgets, regression alerts on PRs, and a quarterly model-cost review.
[VALIDATE IT YOURSELF]
How to check it’s really there.
You do not need to read the code. Ask these questions and demand these artifacts. Vague answers are the finding.
★ Ask your team
- ?What does one request cost us, by intent and by customer?
- ?Would a PR that doubled token usage be caught before it shipped?
- ?When did we last check whether a cheaper model would do?
★ Demand to see
- Per-call cost telemetry attributed to intent and customer
- A cost budget with regression alerts on pull requests
- A recurring model-cost review
● WHAT L0 LOOKS LIKE
The failure mode, in production.
The bill arrives at month end. Nobody knew a template change grew the prompt to 18k tokens. Per-customer compute has exceeded per-customer revenue for three weeks. The fix is rewriting prompt assembly.
Want this layer audited in your stack?
We run the K-Framework against your AI build and hand you the gap list, ranked by what it will cost you in production.