model_pool.json

{
  "schema_version": 10,
  "notes": "Unified 4-model pool for dual-router SWERouterBench benchmarking (UncommonRoute vs semantic-router KNN). IDs taken verbatim from OpenRouter /api/v1/models on 2026-04-22. Exactly one is_high_baseline=true (anthropic/claude-opus-4.6) drives failure-rerun charge. provider='openai_compat' universally (OpenRouter returns OpenAI-compat usage envelopes). cache_control_style='anthropic' only for Claude family so the harness injects ephemeral cache_control blocks; others rely on provider-side cache accounting.",
  "pool": [
    {
      "model_id": "anthropic/claude-opus-4.6",
      "provider": "openai_compat",
      "cache_control_style": "anthropic",
      "is_high_baseline": true
    },
    {
      "model_id": "google/gemini-3-flash-preview",
      "provider": "openai_compat",
      "cache_control_style": null
    },
    {
      "model_id": "minimax/minimax-m2.7",
      "provider": "openai_compat",
      "cache_control_style": null
    },
    {
      "model_id": "deepseek/deepseek-v3.2",
      "provider": "openai_compat",
      "cache_control_style": null
    }
  ]
}