How to Cut AI Costs With Model Routing

How to Cut AI Costs With Model Routing

Cut AI costs by routing work according to the capability it actually needs. Let a small, cheap model handle wake-ups, glue, status checks, and orchestration. Reserve frontier models for building and difficult reasoning. The durable fix is a written routing protocol, not asking operators to make a fresh cost decision every time they send a prompt.

THE CORE IDEA

What is Tiered-Model Cost Discipline?

Tiered-Model Cost Discipline is Vista's framework for assigning each AI task to the least expensive model tier that can complete it reliably. Small models own coordination work. Frontier models own the building and reasoning work where capability changes the quality of the result. The point is not to minimize the cost of every prompt. It is to stop paying for capability a task cannot use.

The distinction fades when a workflow moves quickly. An operator opens the strongest model for a hard problem, gets a good result, and leaves every later prompt in the same thread. Convenience quietly becomes the routing policy.

Tiered-Model Cost Discipline. A Vista framework that routes wake-ups, glue, status prompts, and orchestration to a small, cheap model while reserving frontier models for building and complex reasoning. The objective is reliable work at the right capability tier, not cheap output at any cost.

In the engagements we run, model cost is often blamed on ambitious building tasks. The quieter leak is repetition at the edges: checking, handing off, keeping a process alive, and asking what happened. Those prompts feel too small to govern. Together, they reveal whether the operation has a routing system or only a preferred model.

THE HIDDEN LEAK

Why does AI spend become invisible?

AI spend becomes invisible when prompt habit substitutes for task classification. The operator has a model open, so every request goes there. Nothing breaks loudly. The system simply buys frontier capability for tasks that needed memory, formatting, or a handoff.

Status prompts are the cleanest example. “Keep going,” “check whether the file exists,” and “summarize the last action” do not become better because a frontier model handles them. Neither does moving structured output between tools or waking a paused process. Yet these prompts often inherit the model selected for the hardest step nearby.

The expensive habit is not using a frontier model. It is using one without a reason.

Operators we work with usually see the pattern once they draw the workflow rather than inspect isolated prompts. The large reasoning step is visible and defensible. The chain of small coordination steps around it is diffuse. Tiered routing makes those edges explicit, where they can be assigned deliberately.

THE SETUP

What will you need before you route model work?

You need a task inventory, a plain-language definition of each model tier, an owner for exceptions, and a small set of quality checks. Do not begin with a complicated scoring system. Begin with enough shared language that two operators would route the same task the same way.

What you'll need. A list of recurring AI tasks, access to a small model and a frontier model, a place to record routing defaults, representative test inputs, an acceptance check for important outputs, and a named human who can approve exceptions.

The task list should describe jobs, not departments or tools. “Prepare a decision memo from conflicting evidence” is routable. “Marketing prompt” is not. Good routing begins with the work itself.

THE PROTOCOL

How do you build a model-routing protocol?

Build the protocol in the order below. Each step turns a judgment that currently lives in an operator's head into an operating default that the whole team can inspect.

1. Inventory recurring AI tasks in plain language. Capture the jobs that repeat across a normal workflow, including the tiny prompts between visible deliverables. Name the input, the expected output, and what happens next. Include wake-ups, file checks, formatting, summaries, tool handoffs, drafting, synthesis, and decision support.

   Why it matters: You cannot route work you have not named. The small prompts are easy to omit because no one considers them deliverables, but they are exactly where frontier usage spreads without a capability reason.

2. Classify each task by the capability it requires. Ask whether the job needs retrieval, transformation, coordination, generation, or complex reasoning. Then note the consequence of a weak answer. A task with a simple output and an easy retry belongs in a different class from one that shapes a customer decision or a system design.

   Why it matters: Task labels prevent importance from being confused with difficulty. A message can be important but mechanically simple. A short prompt can require deep judgment. Routing improves when capability and consequence are assessed separately.

3. Set the small model as the default coordination lane. Send wake-ups, keep-going prompts, status checks, structured extraction, formatting, routing decisions, and tool orchestration to the cheap tier. Give it clear schemas and narrow instructions. Escalate only when it encounters an exception the protocol already names.

   Why it matters: Defaults beat willpower. If every operator must remember to downgrade a routine prompt, convenience will eventually win. A structural small-model lane makes cost discipline the normal path through the system.

4. Reserve frontier models for building and difficult reasoning. Route architecture, nuanced synthesis, ambiguous diagnosis, high-stakes drafting, and work that must integrate competing constraints to the stronger tier. State why the frontier tier is required in the task definition, so the exception is legible rather than habitual.

   Why it matters: The goal is not universal cheapness. Frontier capability is valuable when the work can absorb it. Protecting that lane gives demanding tasks the attention they need without letting nearby coordination inherit the same cost.

5. Install acceptance checks and escalation triggers. Define what a passing output looks like for each class. Let simple tasks use deterministic checks such as required fields or a successful handoff. Give complex work a human-in-the-loop gate that reviews reasoning, assumptions, and the decision before it moves forward.

   Why it matters: Routing without evaluation is guesswork. Clear acceptance checks let the small model own routine work confidently, while escalation triggers catch the cases where apparent simplicity hides ambiguity or risk.

6. Review exceptions and update the routing map. Record where the small tier failed, where the frontier tier added no meaningful value, and where operators bypassed the default. Change the task class or instruction when a pattern repeats. Keep one-off surprises as exceptions until they become a real category.

   Why it matters: A routing protocol is an operating system, not a poster. Actual failures show where the classification or guardrail is weak. Review turns those failures into better defaults instead of permanent fear about using the cheaper tier.

This sequence reflects Vista's agent-does-the-work model: the system handles the repeatable operating work, while a person blesses the judgments that deserve human accountability. Model routing is the same division applied inside the AI stack.

THE ROUTING MAP

Which tasks belong in each model tier?

Route by cognitive demand and cost of failure, not by how polished the request sounds. The table is a starting map. Your acceptance checks and escalation rules should decide the boundary in your own workflow.

Task pattern	Default tier	Minimum control
Wake-up, keep-going, or status prompt	Small model	Confirm the next action or current state
Formatting, extraction, or schema conversion	Small model	Validate required fields and output shape
Tool handoff or workflow orchestration	Small model	Confirm the handoff completed and preserve context
Routine summary with clear source material	Small model	Check coverage against the supplied material
New build with interacting requirements	Frontier model	Review the design and test the result
Ambiguous diagnosis or competing evidence	Frontier model	Inspect assumptions and reasoning with a human gate
Decision support with meaningful consequences	Frontier model	Keep the accountable person at the approval point

This map also exposes false economy. A task can appear routine but contain unresolved ambiguity. If a summary must reconcile conflicting evidence, it is no longer simple compression. If orchestration must decide which business rule applies, it may have crossed into reasoning. Route the actual job, not the familiar verb in its prompt.

THE QUALITY CLIFF

When should you refuse the cheaper tier?

Refuse it when the work depends on deep context, multi-step reasoning, novel construction, or a judgment whose error will propagate. The quality cliff on hard work is real. Sending demanding work to a small model can create confident gaps, brittle builds, and a review burden larger than the token savings.

This is the honest counterweight to cost discipline. Cheap models are not a moral preference, and frontier models are not an indulgence. They are different tools. The wrong cheap route produces rework, repeated prompts, and delayed decisions. It can cost more while delivering less.

Across the AI operating work we do, the best signal is not output length. It is how many constraints the task must hold at once. A short architecture choice can demand more capability than a long formatting job. The real-constraint lens keeps the question honest: is model cost actually blocking the workflow, or is weak task definition creating waste at every tier?

THE OPERATING RHYTHM

How do you keep routing discipline from fading?

Put routing ownership into the workflow. Name who can change a default, where exceptions are recorded, and when the map gets reviewed. A protocol that no one owns will decay into personal preference, especially when a deadline makes the strongest model feel like the safest shortcut.

Treat the map as a learning asset. When a small model fails, diagnose the task class, instruction, and acceptance check. When a frontier model handles routine work, ask why the default was bypassed. The answer may reveal a broken handoff.

Teams that want to practice this operating rhythm can use Vista's free AI Lab to work through a real workflow, or go deeper on the full operating pattern inside the AI Cohort. If the harder question is which operator should help design the system, matchmaking for an operator advisor connects the work to someone who can own the implementation with you.

QUESTIONS

Frequently asked questions

What is model routing for AI work?

Model routing is the practice of assigning each AI task to a capability tier before execution. Routine coordination, formatting, and status work defaults to a small model. Building and complex reasoning default to a frontier model. A useful routing system also defines acceptance checks, escalation triggers, and an owner for exceptions.

Should every routine prompt use a small model?

No. Routine appearance is not the same as low cognitive demand. A familiar summary can still require reconciling conflicting evidence, and a tool handoff can contain a meaningful business decision. Use the small tier when the output is bounded, testable, and easy to retry. Escalate when ambiguity or consequence rises.

How do I know when a frontier model is worth using?

Use a frontier model when the work must hold many interacting constraints, reason through ambiguity, construct something novel, or support a consequential decision. The tier earns its place when stronger capability materially changes the result. Record that reason in the routing map so frontier use remains intentional rather than automatic.

Can model routing reduce quality?

Yes, if hard work is mis-routed to a cheaper tier. The likely result is missing reasoning, brittle output, and expensive rework. Quality stays protected when routing includes acceptance checks and escalation triggers. Cost discipline means using the least expensive tier that is reliable, not simply choosing the cheapest available model.

Where should a team start with model routing?

Start with one recurring workflow and inventory every AI task inside it, especially status prompts and handoffs. Classify each task by required capability and consequence of failure. Set a small-model default for coordination, reserve the frontier tier for hard work, then review exceptions after the workflow has run in practice.

NEXT MOVE

Make the routing decision before the prompt

The durable savings come from removing the repeated choice. This week, map one workflow and mark every task as coordination or building and reasoning. Set the small tier as the coordination default, name the escalation trigger, and protect the frontier lane for work that can use it. That is Tiered-Model Cost Discipline in operation.