{
  "runs": [
    {
      "id": "samaya-system",
      "run_name": "samaya-system",
      "label": "samaya-system",
      "system": "samaya-composite",
      "models": [
        "samaya"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.508272,
      "accuracy_essential": null,
      "cost_per_query": 0.9656,
      "latency_per_query": 218.1,
      "total_cost": 212.44,
      "has_per_query": false
    },
    {
      "id": "fa-v2-opus-4-8",
      "run_name": "fa-v2-opus-4-8",
      "label": "fa-v2-opus-4-8",
      "system": "finance-agent-v2",
      "models": [
        "claude-opus-4-8"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.450399,
      "accuracy_essential": null,
      "cost_per_query": 2.6109,
      "latency_per_query": 155.8,
      "total_cost": 574.4,
      "has_per_query": false
    },
    {
      "id": "fa-v2-fable-5",
      "run_name": "fa-v2-fable-5",
      "label": "fa-v2-fable-5",
      "system": "finance-agent-v2",
      "models": [
        "claude-fable-5"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.491922,
      "accuracy_essential": null,
      "cost_per_query": 4.0592,
      "latency_per_query": 164.5,
      "total_cost": 893.02,
      "has_per_query": false
    },
    {
      "id": "fa-v2-kimi-k2-6",
      "run_name": "fa-v2-kimi-k2-6",
      "label": "fa-v2-kimi-k2-6",
      "system": "finance-agent-v2",
      "models": [
        "kimi-k2.6"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.323175,
      "accuracy_essential": null,
      "cost_per_query": 0.9487,
      "latency_per_query": 201.1,
      "total_cost": 208.72,
      "has_per_query": false
    },
    {
      "id": "fa-v2-glm-5-2",
      "run_name": "fa-v2-glm-5-2",
      "label": "fa-v2-glm-5-2",
      "system": "finance-agent-v2",
      "models": [
        "glm-5.2"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.324786,
      "accuracy_essential": null,
      "cost_per_query": 0.8243,
      "latency_per_query": 203.3,
      "total_cost": 181.36,
      "has_per_query": false
    },
    {
      "id": "fa-v2-deepseek-v4-pro",
      "run_name": "fa-v2-deepseek-v4-pro",
      "label": "fa-v2-deepseek-v4-pro",
      "system": "finance-agent-v2",
      "models": [
        "deepseek-v4-pro"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.405397,
      "accuracy_essential": null,
      "cost_per_query": 0.677,
      "latency_per_query": 202.6,
      "total_cost": 148.93,
      "has_per_query": false
    },
    {
      "id": "fa-v2-gpt-5-5",
      "run_name": "fa-v2-gpt-5-5",
      "label": "fa-v2-gpt-5-5",
      "system": "finance-agent-v2",
      "models": [
        "gpt-5.5"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.434736,
      "accuracy_essential": null,
      "cost_per_query": 2.8005,
      "latency_per_query": 233,
      "total_cost": 616.11,
      "has_per_query": false
    },
    {
      "id": "fa-v2-gpt-5-6-sol",
      "run_name": "fa-v2-gpt-5-6-sol",
      "label": "fa-v2-gpt-5-6-sol",
      "system": "finance-agent-v2",
      "models": [
        "gpt-5.6-sol"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.468300,
      "accuracy_essential": null,
      "cost_per_query": 3.0324,
      "latency_per_query": 170.7,
      "total_cost": 667.13,
      "has_per_query": false
    },
    {
      "id": "fa-v2-gemini-3-1-pro",
      "run_name": "fa-v2-gemini-3-1-pro",
      "label": "fa-v2-gemini-3-1-pro",
      "system": "finance-agent-v2",
      "models": [
        "gemini-3.1-pro"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.305132,
      "accuracy_essential": null,
      "cost_per_query": 1.7228,
      "latency_per_query": 221.9,
      "total_cost": 379.01,
      "has_per_query": false
    },
    {
      "id": "web-search-gpt-5-5",
      "run_name": "web-search-gpt-5-5",
      "label": "web-search-gpt-5-5",
      "system": "web-search",
      "models": [
        "gpt-5.5"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.207298,
      "accuracy_essential": null,
      "cost_per_query": 0.5464,
      "latency_per_query": 192.1,
      "total_cost": 120.21,
      "has_per_query": false
    },
    {
      "id": "web-search-opus-4-8",
      "run_name": "web-search-opus-4-8",
      "label": "web-search-opus-4-8",
      "system": "web-search",
      "models": [
        "claude-opus-4-8"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.329841,
      "accuracy_essential": null,
      "cost_per_query": 1.4737,
      "latency_per_query": 53.9,
      "total_cost": 324.22,
      "has_per_query": false
    },
    {
      "id": "web-search-gemini-3-1-pro",
      "run_name": "web-search-gemini-3-1-pro",
      "label": "web-search-gemini-3-1-pro",
      "system": "web-search",
      "models": [
        "gemini-3.1-pro"
      ],
      "benchmark_mode": false,
      "predefined_mode": null,
      "num_records": 220,
      "accuracy": 0.307154,
      "accuracy_essential": null,
      "cost_per_query": 0.2406,
      "latency_per_query": 91.1,
      "total_cost": 52.94,
      "has_per_query": false
    }
  ]
}