{
  "version": "0.2.0",
  "updated": "2026-06-27",
  "routing_policies": [
    {
      "id": "cheapest_safe",
      "name": "cheapest_safe",
      "description": "Use low-cost models for simple extraction/classification; escalate when confidence is low.",
      "use_when": ["simple extraction or classification", "high-volume batch jobs", "cost is the primary constraint", "loose latency tolerance"],
      "default_providers": ["low_cost_text", "router_text"],
      "fallback": "premium_text",
      "markup": "provider_cost + 12% routing fee; lane targets <$0.50/M tokens"
    },
    {
      "id": "premium_reasoning",
      "name": "premium_reasoning",
      "description": "Use direct premium APIs for complex coding, legal-style analysis, or high-trust workflows.",
      "use_when": ["complex reasoning or coding", "legal/financial-grade analysis", "customer-facing output", "long-context synthesis"],
      "default_providers": ["premium_text"],
      "fallback": "router_text",
      "markup": "provider_cost + 12% routing fee; lane runs $2-15/M tokens"
    },
    {
      "id": "media_low_cost",
      "name": "media_low_cost",
      "description": "Use lower-cost image/video providers when the quality target permits; retry on a premium provider when output is rejected.",
      "use_when": ["draft or iterative media", "cost-sensitive generation", "high volume"],
      "default_providers": ["media"],
      "fallback": "media",
      "markup": "provider_cost + media margin"
    },
    {
      "id": "byok",
      "name": "byok",
      "description": "Customer brings their own provider key; ToolCall Store charges orchestration, logging, and safety fees only.",
      "use_when": ["customer has direct provider contracts", "data-residency or compliance requirements", "highest cost control"],
      "default_providers": ["premium_text", "router_text", "low_cost_text", "media"],
      "fallback": "router_text",
      "markup": "orchestration/safety fee only (no provider markup)"
    }
  ],
  "provider_groups": {
    "premium_text": [
      { "name": "OpenAI", "example_models": ["GPT-5.5", "GPT-5-mini"], "notes": "Direct API; strong general reasoning and tool use." },
      { "name": "Anthropic", "example_models": ["Claude Opus 4.5", "Claude Sonnet 4.5", "Claude Haiku 4.5"], "notes": "Direct API; strong long-context and coding; Haiku for cheaper high-quality." },
      { "name": "Google Gemini", "example_models": ["Gemini 2.5 Pro", "Gemini 2.5 Flash"], "notes": "Direct API; very large context; Flash is a cheap fast tier." }
    ],
    "router_text": [
      { "name": "OpenRouter", "example_models": ["aggregates 300+ models via one key"], "notes": "Single key, broad failover, pass-through pricing." },
      { "name": "Together AI", "example_models": ["Llama", "Qwen", "DeepSeek (open weights)"], "notes": "Hosted open-weight models at competitive rates." }
    ],
    "low_cost_text": [
      { "name": "DeepSeek", "example_models": ["DeepSeek-V3.2", "DeepSeek-R1"], "notes": "Very low cost; R1 for cheap reasoning." },
      { "name": "Groq", "example_models": ["Llama 3.3 70B"], "notes": "Ultra-fast LPU inference; great for latency-sensitive low-cost calls." },
      { "name": "Cloudflare Workers AI", "example_models": ["edge-hosted open models"], "notes": "Edge inference, low latency, low cost." },
      { "name": "Fireworks AI", "example_models": ["open-weight models, fast"], "notes": "Fast hosted open-weight inference." },
      { "name": "Mistral", "example_models": ["Mistral Large", "Mistral Small"], "notes": "Strong European option; Small is cheap." },
      { "name": "Cerebras", "example_models": ["ultra-fast Llama inference"], "notes": "Wafer-scale inference; extremely high tokens/sec." }
    ],
    "media": [
      { "name": "Leonardo.Ai", "example_models": ["image"], "notes": "Image generation; user credit may already be available." },
      { "name": "Gemini Image", "example_models": ["Imagen", "Nano Banana"], "notes": "Google image generation/editing." },
      { "name": "fal.ai", "example_models": ["FLUX", "video models"], "notes": "Fast image and video; good for FLUX and short video." },
      { "name": "Replicate", "example_models": ["broad model hosting"], "notes": "Wide catalog of community image/video models." },
      { "name": "Runware", "example_models": ["low-cost image"], "notes": "Low-cost high-throughput image generation." },
      { "name": "Stability AI", "example_models": ["Stable Diffusion 3.5"], "notes": "Direct SD models." }
    ]
  },
  "pricing_guidance": {
    "note": "Exact token and media prices vary by provider and change frequently. ToolCall Store pulls live provider price tables at quote time.",
    "low_cost_lane_target": "<$0.50 per 1M tokens",
    "premium_lane_range": "$2-15 per 1M tokens",
    "recommendation": "Agents should request a quote (/v1/quotes) rather than assume prices; respect max_cost_usd on each call."
  }
}