Tiering is the discipline, not cheaper models
Most agentic systems route every task to the most capable model available. That is not a strategy — it is the absence of one, and it converts a solvable engineering question into an unbounded operating expense.
On our own reference system the answer turned out to be five percent: only synthesis, verification of other agents’ work, and decisions that move capital reach the frontier tier. The result is a research organisation operating at $0.88 per terminal verdict against an industry norm of analyst-months.
Runaway cost is the number one reason agent projects get cancelled.
Gartner puts it first on the list of three reasons it expects 40% of agent projects to be cancelled by 2027: agents run continuously, generating API calls and consuming tokens around the clock, accumulating cost far beyond the original estimate.
Most agentic systems route every task to the most capable model available. That is not a strategy — it is the absence of one.
- The reflex is defensible. Frontier models are better at everything, so using them everywhere looks like the safe choice.
- It hides the real question: which tasks actually require judgement? On our own system the answer was five percent.
- Spend arrives as one number. Nobody can tie a dollar to a decision, so nobody can cut cost without cutting capability.
- Agents do not sleep. A loop that costs pennies in testing runs continuously in production.
- Tiered model routing. Every feature declares the model class its work requires; the runtime enforces it.
- Hard budget ceilings. Per-task, per-agent and per-tenant budgets that throw rather than silently degrade.
- Per-call ledgering. Model, token count, cost and prompt hash recorded for every single call.
- Mismatch detection. Intended model versus actual model compared continuously; every divergence flagged.
- Anomaly alerting tied to output, not raw volume — so a busy day does not read as an incident.
Five tiers. The frontier one handles five percent.
This is the actual routing table from our reference system, with measured share of calls. The saving does not come from using worse models — it comes from being specific about where judgement is genuinely required.
| Tier | Model class | Assigned work | Share of calls |
|---|---|---|---|
| T0 | no model | Polling, status checks, test runs, shipping | — |
| T1 | light | Fetch, format, classify against explicit criteria — never permitted to issue a verdict | — |
| T2 | mid | The default workforce: coding, tests, refactors, research passes | 68% |
| T3 | heavy | Design and risk logic: multi-file architecture, gate logic, cross-module debugging | 27% |
| T4 | frontier | Synthesis, verification of other agents’ output, and any decision that moves capital | 5% |
$0.88 per terminal research verdict.
Against an industry norm measured in analyst-months — roughly $10,000 to $100,000+ of loaded cost for the same output. The figures below are from our own reference system, currently in paper trading.
Tiered model routing
Work is matched to the cheapest tier that can do it. Polling uses no model at all; verification is never the step that gets cheapened.
Hard budget ceilings
Per-stage token budgets enforced in code. A run that exceeds its ceiling parks for human review instead of quietly billing.
Per-call ledgering
Model, tokens, cost and prompt hash recorded for every call. Spend becomes a queryable table, not a monthly surprise.