Model Tier Enforcement

Organize models into cost tiers and restrict agents to a maximum tier. When an agent requests a model above its allowed tier, TealEngine automatically downgrades to the highest permitted model.

Policy JSON

{
  "cost": {
    "model_tiers": {
      "tier_1": ["gpt-4o", "claude-3-opus"],
      "tier_2": ["gpt-4o-mini", "claude-3-sonnet"],
      "tier_3": ["gpt-3.5-turbo", "claude-3-haiku"]
    },
    "max_tier": 2,
    "action_on_violation": "DEGRADE",
    "downgrade_to_tier": 2
  }
}

policy = {
    "cost": {
        "model_tiers": {
            "tier_1": ["gpt-4o", "claude-3-opus"],
            "tier_2": ["gpt-4o-mini", "claude-3-sonnet"],
            "tier_3": ["gpt-3.5-turbo", "claude-3-haiku"],
        },
        "max_tier": 2,
        "action_on_violation": "DEGRADE",
        "downgrade_to_tier": 2,
    }
}

What It Does

Defines model tiers from most expensive (tier 1) to least expensive (tier 3)
Enforces a maximum tier — agents cannot use models above the configured tier
Automatically downgrades requests for tier 1 models to the best available tier 2 model
Prevents cost surprises from agents that autonomously select expensive models

Example

Agent requests gpt-4o (tier 1) when max tier is 2: Decision object:

{
  "action": "DEGRADE",
  "risk_score": 0.5,
  "reasons": ["model_tier_violation"],
  "evidence": {
    "requested_model": "gpt-4o",
    "requested_tier": 1,
    "max_tier": 2,
    "downgraded_to": "gpt-4o-mini",
    "downgraded_tier": 2
  }
}

Install

# Install via CLI
tealtiger policy install model-tier-enforcement

# Or copy the JSON above into your tealtiger.config.json

Customize

Setting	Default	Description
`model_tiers`	See above	Map of tier names to model arrays. Add your own tiers and models
`max_tier`	`2`	Maximum allowed tier (1 = most expensive, 3 = cheapest)
`action_on_violation`	`DEGRADE`	Action on tier violation. Options: `DEGRADE`, `DENY`, `MONITOR`
`downgrade_to_tier`	`2`	Target tier for automatic downgrade

Use DENY instead of DEGRADE when model capability matters more than availability. For example, a code generation agent may produce incorrect results with a cheaper model — better to deny and alert than silently downgrade.

Budget — Daily $10 — Combine tier enforcement with spending limits
Model Allowlist — Restrict which models are available at all
Fallback Chain — Automatic failover across providers
Cost Tracking Documentation — Full API reference

Getting Started

Core Concepts

Governance Domains

Policy Authoring

Architecture

Integrations

Deployment

API Reference

Cookbook

Compliance & GRC

Guides

Reference

Versions

Playground

Dashboard

Blog

Policy JSON

What It Does

Example

Install

Customize

​Policy JSON

​What It Does

​Example

​Install

​Customize

​Related

Policy JSON

What It Does

Example

Install

Customize

Related