{
  "version": 1,
  "tiers": {
    "nano": {
      "label": "GPT-5.4 nano",
      "modelId": "gpt-5.4-nano",
      "provider": "openai",
      "taskTypes": ["classification", "extraction", "ranking", "labeling", "summarization"],
      "maxContextTokens": 32000,
      "costMultiplier": 0.1
    },
    "mini": {
      "label": "GPT-5.4 mini",
      "modelId": "gpt-5.4-mini",
      "provider": "openai",
      "taskTypes": ["code-edit", "test-generation", "review", "tool-use", "debugging"],
      "maxContextTokens": 200000,
      "costMultiplier": 0.4
    },
    "frontier": {
      "label": "GPT-5.5",
      "modelId": "gpt-5.5",
      "provider": "openai",
      "taskTypes": ["architecture", "cross-file", "complex-debugging", "large-context"],
      "maxContextTokens": 1000000,
      "costMultiplier": 1.0,
      "budgetDefaults": {
        "tokenCap": 500000,
        "requireReason": true
      }
    },
    "localFrontier": {
      "label": "Configured local frontier",
      "modelId": null,
      "provider": "openai-compatible",
      "taskTypes": [],
      "maxContextTokens": 1000000,
      "costMultiplier": 0.0,
      "budgetDefaults": {
        "tokenCap": 2000000,
        "requireReason": false
      },
      "notes": "Self-hosted open-source tier. The runtime derives the served model from THUMBGATE_MODEL_ROLE_NORMAL, THUMBGATE_LOCAL_MODEL, THUMBGATE_MODEL_ID, or THUMBGATE_LOCAL_MODEL_FAMILY. Zero marginal API cost; no token budget enforcement needed."
    }
  },
  "escalationRules": {
    "contextThreshold": 200000,
    "failureRetryThreshold": 2,
    "highRiskEscalation": true,
    "architectureTags": ["architecture", "cross-file", "multi-module", "system-design"]
  }
}