{
  "models": [
    "SWE-2",
    "SWE-1.7",
    "GPT-5.6 Sol",
    "Fable 5",
    "GPT-6 Astra",
    "Grok 4.6",
    "Kimi K3"
  ],
  "colors": {
    "SWE-2": "#2200ff",
    "SWE-1.7": "#2200ff",
    "GPT-5.6 Sol": "#10a37f",
    "Fable 5": "#d97757",
    "GPT-6 Astra": "#10a37f",
    "Grok 4.6": "#5c616a",
    "Kimi K3": "#7c3aed"
  },
  "lab_colors": {},
  "harness": {
    "SWE-2": "chisel",
    "SWE-1.7": "chisel",
    "GPT-5.6 Sol": "mini-swe-agent",
    "Fable 5": "mini-swe-agent",
    "GPT-6 Astra": "mini-swe-agent",
    "Grok 4.6": "mini-swe-agent",
    "Kimi K3": "mini-swe-agent"
  },
  "efforts": {
    "SWE-2": [
      "medium",
      "high",
      "max"
    ],
    "SWE-1.7": [
      "max"
    ],
    "GPT-5.6 Sol": [
      "low",
      "medium",
      "high",
      "xhigh",
      "max"
    ],
    "Fable 5": [
      "low",
      "medium",
      "high",
      "xhigh"
    ],
    "GPT-6 Astra": [
      "low",
      "medium",
      "high",
      "xhigh",
      "max"
    ],
    "Grok 4.6": [
      "low",
      "medium",
      "high",
      "xhigh"
    ],
    "Kimi K3": [
      "max"
    ]
  },
  "subsets": {
    "main": 113
  },
  "data": {
    "SWE-2": {
      "medium": {
        "main": {
          "new_score": 0.6430678466076696,
          "cost": 0.45198814922566355,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "high": {
        "main": {
          "new_score": 0.69,
          "cost": 0.9981015987831859,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "max": {
        "main": {
          "new_score": 0.73,
          "cost": 1.3426005914601769,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "SWE-1.7": {
      "max": {
        "main": {
          "new_score": 0.3769911504424779,
          "cost": 2.7731428317407407,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "GPT-5.6 Sol": {
      "low": {
        "main": {
          "new_score": 0.45353982300884954,
          "cost": 1.0742664646017699,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "medium": {
        "main": {
          "new_score": 0.6106194690265486,
          "cost": 1.8620318075221238,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "high": {
        "main": {
          "new_score": 0.6940133037694013,
          "cost": 3.4698331108647453,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "xhigh": {
        "main": {
          "new_score": 0.7073170731707317,
          "cost": 4.703663860310422,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "max": {
        "main": {
          "new_score": 0.7266666666666667,
          "cost": 8.386436346666667,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "Fable 5": {
      "low": {
        "main": {
          "new_score": 0.5958429561200924,
          "cost": 3.757873125874126,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "medium": {
        "main": {
          "new_score": 0.6536697247706422,
          "cost": 6.088186528935186,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "high": {
        "main": {
          "new_score": 0.686046511627907,
          "cost": 9.177638214788733,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "xhigh": {
        "main": {
          "new_score": 0.6991150442477876,
          "cost": 13.414521495535714,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "GPT-6 Astra": {
      "low": {
        "main": {
          "new_score": 0.6703539823008849,
          "cost": 2.188762706637168,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "medium": {
        "main": {
          "new_score": 0.7278761061946902,
          "cost": 4.379527705752213,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "high": {
        "main": {
          "new_score": 0.7323008849557522,
          "cost": 5.723721956194691,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "xhigh": {
        "main": {
          "new_score": 0.7411504424778761,
          "cost": 6.52377356460177,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "max": {
        "main": {
          "new_score": 0.7323008849557522,
          "cost": 12.369032401327434,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "Grok 4.6": {
      "low": {
        "main": {
          "new_score": 0.41648106904231624,
          "cost": 1.0423604231625836,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "medium": {
        "main": {
          "new_score": 0.6747787610619469,
          "cost": 3.4489837522123894,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "high": {
        "main": {
          "new_score": 0.6518847006651884,
          "cost": 4.3848732239467845,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      },
      "xhigh": {
        "main": {
          "new_score": 0.6674057649667405,
          "cost": 5.497668745011087,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    },
    "Kimi K3": {
      "max": {
        "main": {
          "new_score": 0.6851441241685144,
          "cost": 4.654682129933482,
          "correct": null,
          "tokens": null,
          "duration_min": null,
          "tool_calls": null,
          "ote": null
        }
      }
    }
  }
}