{
  "created_at": "2026-07-09T07:49:36.588432+00:00",
  "source_data": "quantbench-all-models-data-with-grok45.json",
  "source_artifact": "x-ai__grok-4.5__strict__20260709T073928Z-c5fdf7a96e.json",
  "outputs": {
    "master_png": "grok-4.5-top5-quantbench-master-1600.png",
    "social_png": "grok-4.5-top5-quantbench-1080.png",
    "social_jpg": "grok-4.5-top5-quantbench-1080.jpg"
  },
  "canvas": {
    "master": [
      1600,
      1600
    ],
    "social": [
      1080,
      1080
    ]
  },
  "top5": [
    {
      "label": "GPT-5.5",
      "model": "openai/gpt-5.5",
      "score": 0.9653535382668413,
      "json": 1.0,
      "action": 0.9811320754716981,
      "false": 0,
      "missed": 0,
      "cost": 2.0262250000000006,
      "tokens": 150120,
      "run_id": "20260708T051356Z-ef5567941d"
    },
    {
      "label": "Qwen 3.7 Max",
      "model": "qwen/qwen3.7-max",
      "score": 0.964055420076546,
      "json": 1.0,
      "action": 0.9905660377358491,
      "false": 0,
      "missed": 0,
      "cost": 0.71436375,
      "tokens": 266903,
      "run_id": "20260708T051356Z-c16015de07"
    },
    {
      "label": "Qwen 3.6 27B",
      "model": "qwen/qwen3.6-27b",
      "score": 0.9620379153000008,
      "json": 1.0,
      "action": 0.9716981132075472,
      "false": 0,
      "missed": 0,
      "cost": 0.76180498,
      "tokens": 416720,
      "run_id": "20260708T052944Z-a64e377805"
    },
    {
      "label": "Grok 4.5",
      "model": "x-ai/grok-4.5",
      "score": 0.9582440047366646,
      "json": 1.0,
      "action": 0.9716981132075472,
      "false": 0,
      "missed": 0,
      "cost": 1.131042,
      "tokens": 277753,
      "run_id": "20260709T073928Z-c5fdf7a96e"
    },
    {
      "label": "GLM 5.2",
      "model": "z-ai/glm-5.2",
      "score": 0.9405168084374604,
      "json": 1.0,
      "action": 0.9339622641509434,
      "false": 0,
      "missed": 5,
      "cost": 0.47946947999999995,
      "tokens": 156742,
      "run_id": "20260707T141357Z-6a71127d3f"
    }
  ],
  "grok45_emphasis": {
    "rank": 4,
    "score": 0.9582440047366646,
    "delta_score_points_vs_grok43": 3.522724941096733,
    "delta_action_points_vs_grok43": 5.660377358490565
  },
  "leaderboard_manifest": [
    {
      "run_id": "20260708T051356Z-ef5567941d",
      "artifact_id": "openai__gpt-5.5__strict__20260708T051356Z-ef5567941d.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T051356Z-c16015de07",
      "artifact_id": "qwen__qwen3.7-max__strict__20260708T051356Z-c16015de07.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T052944Z-a64e377805",
      "artifact_id": "qwen__qwen3.6-27b__strict__20260708T052944Z-a64e377805.json",
      "source_group": "doc-batch-20260708-qwen36-max8000",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260709T073928Z-c5fdf7a96e",
      "artifact_id": "x-ai__grok-4.5__strict__20260709T073928Z-c5fdf7a96e.json",
      "source_group": "grok-4.5-20260709",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T141357Z-6a71127d3f",
      "artifact_id": "glm-5.2.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T141357Z-3b6fce3999",
      "artifact_id": "deepseek-v4-pro.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T161318Z-c8f1f5311b",
      "artifact_id": "deepseek__deepseek-v4-flash__strict__20260707T161318Z-c8f1f5311b.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T051356Z-a89ccb8d9f",
      "artifact_id": "minimax__minimax-m3__strict__20260708T051356Z-a89ccb8d9f.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T051356Z-e2a3724944",
      "artifact_id": "x-ai__grok-4.3__strict__20260708T051356Z-e2a3724944.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T130202Z-a4347f530c",
      "artifact_id": "20260707T130202Z-a4347f530c.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T161318Z-08e27cc1bc",
      "artifact_id": "anthropic__claude-opus-4.8__strict__20260707T161318Z-08e27cc1bc.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T051356Z-b1279adf92",
      "artifact_id": "moonshotai__kimi-k2.7-code__strict__20260708T051356Z-b1279adf92.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T113837Z-c948c50b94",
      "artifact_id": "metrics.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260707T161318Z-9577a732ac",
      "artifact_id": "tencent__hy3_free__strict__20260707T161318Z-9577a732ac.json",
      "source_group": "expanded-baselines-20260707",
      "prompt_mode": "strict",
      "repeats": 1
    },
    {
      "run_id": "20260708T051356Z-482cef79dd",
      "artifact_id": "tencent__hy3__strict__20260708T051356Z-482cef79dd.json",
      "source_group": "doc-batch-20260708",
      "prompt_mode": "strict",
      "repeats": 1
    }
  ]
}