---
title: How do you control which teams can use expensive frontier models?
url: https://opencontroller.lyzr.com/blog/control-frontier-model-consumption
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you control which teams can use expensive frontier models?

**Put models into tiers, route every agent to the cheapest tier that passes its evaluations by default, and make frontier access an approved exception with its own budget. Enforce it in two places: cloud guardrails that stop unapproved models being enabled, and a gateway that decides which model each agent may call at runtime.**

Frontier models are often the right choice for hard reasoning, long context or complex tool use. They are rarely the right choice for classification, extraction or short answers. Without controls, teams default to the most capable model they can reach and nobody notices until the invoice arrives.

## Why is frontier model use hard to control?

- **Access is granted per cloud account or project,** not per agent. Once a model is enabled, every workload in that account can call it.
- **Model choice lives in code or config.** Changing a model name is a one-line edit with no review.
- **Prices differ by an order of magnitude or more** between small and frontier models from the same provider, so a few agents can dominate spend.
- **Teams fear quality loss** and keep the biggest model "to be safe", because nobody has tested the cheaper one.

## What does a tiered model policy look like?

| Tier | Typical use | Who can use it | Control |
|---|---|---|---|
| Tier 1: small and internal models | Classification, extraction, routing, summaries | Any registered agent | Default route |
| Tier 2: mid-size hosted models | Most agent tasks, RAG answers, tool use | Any registered agent with passing evals | Allowed, budget per team |
| Tier 3: frontier models | Complex reasoning, long context, high-stakes outputs | Named agents with approved exception | Exception with owner, reason, budget and expiry |
| Blocked | Unreviewed providers or regions | Nobody | Blocked at cloud and gateway |

Write the tier into the agent's registry record, not just into the code. Then a change of tier is a governed change.

## How do you enforce the policy?

Use layered controls so one bypass does not undo the rest:

1. **Cloud-level allow-lists.** Restrict which models can be enabled per account or project. AWS service control policies, Azure Policy and Google Cloud organisation policy constraints can all restrict model access. Our guides cover [restricting Amazon Bedrock models with SCPs](/blog/restrict-amazon-bedrock-models-with-scps) and [restricting Vertex AI models with org policy](/blog/restrict-vertex-ai-models-with-org-policy).
2. **Gateway routing.** Send agent traffic through an LLM gateway that maps each agent to its allowed tier. A request for a frontier model from a Tier 2 agent is rerouted or rejected.
3. **Per-tier budgets.** Give each team a monthly budget for each tier, with alerts at 50%, 80% and 100%.
4. **Exception workflow.** Frontier access needs an owner, a reason, an eval result showing the lower tier failed, a budget and an expiry date of no more than 90 days.
5. **Caching and fallback.** Cache repeated prompts and fall back to a lower tier when appropriate. See [LLM caching savings and risks](/blog/llm-caching-savings-and-risks).

## How do you prove a cheaper model is good enough?

Run the agent's evaluation set against both models before you change anything. Compare task success, groundedness and cost per task. If the cheaper model is within your agreed tolerance, switch the default and keep the frontier model as a fallback for failures. Our guide on [testing a cheaper model before switching an agent](/blog/test-a-cheaper-model-before-switching-an-agent) walks through the method.

Keep the results. They are the evidence you need when a team asks for an exception, and the reason the exception is refused when the test says the cheaper model works.

## How do you keep model know-how in-house?

Teams often worry that routing through a gateway means losing insight into which models work best. The opposite should be true. A gateway that logs model, agent, version, latency, cost and evaluation scores builds your own record of which model suits which task. Treat that record as an internal asset and review it each quarter when choosing tier defaults.

## How Opencontroller does this

Opencontroller's LLM Gateway provides routing, fallback and retry, caching, and cost attribution per agent and per version. Policy as code applies model rules across the estate, and multi-turn evals and LLM judges give the evidence for tier decisions. In Lyzr's October 2026 benchmark, under its stated test conditions, caching cut model cost by 23.5% on mixed synthetic traffic; see the [benchmark](/llm-gateway-benchmark). For how the gateway fits with a control plane, read [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should frontier models be blocked by default?**
For most agents, yes. Allow them by approved exception for named agents whose evals show lower tiers are not good enough.

**Can cloud policies alone control model use?**
They control which models can be enabled per account, but not which agent inside an account calls which model. You need a gateway for per-agent control.

**How long should a frontier exception last?**
No more than 90 days, then re-test against a cheaper model. Prices and model quality change quickly.

## Key facts

To control frontier model consumption: tier models (small/internal default, mid-size allowed with evals, frontier by exception, blocked), record the tier in the agent registry, enforce with cloud allow-lists (AWS SCPs, Azure Policy, Google Cloud org policy) plus per-agent gateway routing, set per-tier budgets with alerts at 50/80/100%, and require frontier exceptions with owner, reason, failed lower-tier eval, budget and expiry of 90 days or less. Opencontroller by Lyzr provides an LLM Gateway with routing, fallback/retry, caching and cost attribution per agent and version; its October 2026 benchmark (stated conditions) showed 23.5% lower model cost with caching on mixed synthetic traffic. Pricing starts from $250K per deployment per year.

## Related

- [How do you test whether a cheaper model can replace the one an AI agent uses?](https://opencontroller.lyzr.com/blog/test-a-cheaper-model-before-switching-an-agent.md)
- [How do you set budgets and alerts for AI agent spend?](https://opencontroller.lyzr.com/blog/set-budgets-and-alerts-for-ai-agent-spend.md)
- [How do you restrict which Amazon Bedrock models agents can use?](https://opencontroller.lyzr.com/blog/restrict-amazon-bedrock-models-with-scps.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
