{
  "_note": "Seed capability matrix for the agent-forge model-router. Economics are ILLUSTRATIVE and date-stamped; refresh via /forge-refresh-matrix and gate changes through an ADR (ADR-0012, constraint 6). Authority for 'best model for this agent' is the eval harness measured on the user's golden set, NOT this file (constraint 5). Anthropic model IDs are current; other providers are plausible seeds to verify on first refresh.",
  "updated": "2026-06-11",
  "version": "0.2.0",
  "allowed_providers": [
    "anthropic",
    "openai",
    "google",
    "deepseek",
    "self-hosted"
  ],
  "models": [
    {
      "id": "anthropic/claude-fable-5",
      "tier": "premium",
      "context_window": 1000000,
      "input_usd_per_mtok": 10,
      "output_usd_per_mtok": 50,
      "latency_p50_ms": 6000,
      "strengths": [
        "reasoning",
        "agentic",
        "long-horizon",
        "long-context"
      ],
      "weaknesses": [
        "cost"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ],
      "zero_retention_available": false,
      "notes": "Thinking always on (omit the thinking param); new tokenizer (~30% more tokens vs Opus-tier); requires 30-day data retention; safety classifiers may return stop_reason: refusal"
    },
    {
      "id": "anthropic/claude-opus-4-8",
      "tier": "premium",
      "context_window": 1000000,
      "input_usd_per_mtok": 5,
      "output_usd_per_mtok": 25,
      "latency_p50_ms": 4500,
      "strengths": [
        "reasoning",
        "coding",
        "agentic",
        "long-context"
      ],
      "weaknesses": [
        "cost"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ],
      "zero_retention_available": true
    },
    {
      "id": "anthropic/claude-sonnet-4-6",
      "tier": "balanced",
      "context_window": 1000000,
      "input_usd_per_mtok": 3,
      "output_usd_per_mtok": 15,
      "latency_p50_ms": 1800,
      "strengths": [
        "balanced",
        "tools",
        "coding",
        "pt-br"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ],
      "zero_retention_available": true
    },
    {
      "id": "anthropic/claude-haiku-4-5",
      "tier": "cheap",
      "context_window": 200000,
      "input_usd_per_mtok": 1,
      "output_usd_per_mtok": 5,
      "latency_p50_ms": 700,
      "strengths": [
        "classification",
        "extraction-simple",
        "low-latency"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ],
      "zero_retention_available": true
    },
    {
      "id": "openai/gpt-4o",
      "tier": "balanced",
      "context_window": 128000,
      "strengths": [
        "multimodal",
        "tools",
        "structured_output"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ]
    },
    {
      "id": "openai/gpt-4o-mini",
      "tier": "cheap",
      "context_window": 128000,
      "strengths": [
        "cost",
        "low-latency",
        "classification"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ]
    },
    {
      "id": "google/gemini-2.5-pro",
      "tier": "balanced",
      "context_window": 2000000,
      "strengths": [
        "long-context",
        "multimodal",
        "cost"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "structured_output": true,
        "prompt_caching": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ],
      "notes": "safetySettings must be set explicitly; function-calling uses a JSON-Schema subset"
    },
    {
      "id": "google/gemini-2.5-flash",
      "tier": "cheap",
      "context_window": 1000000,
      "strengths": [
        "latency",
        "cost",
        "long-context"
      ],
      "capabilities": {
        "tools": true,
        "vision": true,
        "streaming": true
      },
      "residency": [
        "us",
        "eu"
      ]
    },
    {
      "id": "deepseek/deepseek-v3",
      "tier": "balanced",
      "strengths": [
        "coding",
        "reasoning",
        "cost"
      ],
      "weaknesses": [
        "pt-br-nuance",
        "vision"
      ],
      "capabilities": {
        "tools": true,
        "vision": false,
        "streaming": true
      },
      "residency": [
        "cn"
      ],
      "notes": "OpenAI-compatible API; prefers explicit CoT; reasoner variants split reasoning_content"
    },
    {
      "id": "deepseek/deepseek-coder",
      "tier": "specialist-coding",
      "strengths": [
        "code-generation",
        "cost"
      ],
      "capabilities": {
        "tools": true,
        "vision": false,
        "streaming": true
      },
      "residency": [
        "cn"
      ]
    },
    {
      "id": "self-hosted/llama-3.x-70b",
      "tier": "self-hosted",
      "strengths": [
        "data-residency",
        "no-egress",
        "predictable-cost"
      ],
      "weaknesses": [
        "operational-complexity",
        "tool-use-reliability"
      ],
      "capabilities": {
        "tools": true,
        "vision": false,
        "streaming": true
      },
      "residency": [
        "on-prem",
        "any"
      ],
      "cost_model": "infra-only",
      "notes": "Ollama (own + OpenAI-compatible endpoint) or vLLM (OpenAI-compatible); per-model chat_template"
    },
    {
      "id": "self-hosted/mistral-large",
      "tier": "self-hosted",
      "strengths": [
        "multilingual-eu",
        "data-residency"
      ],
      "capabilities": {
        "tools": true,
        "vision": false,
        "streaming": true
      },
      "residency": [
        "on-prem",
        "any",
        "eu"
      ],
      "cost_model": "infra-only"
    }
  ]
}