[
  {
    "id": "coding-assistant",
    "name": "Coding Assistant",
    "emoji": "\ud83d\udc68\u200d\ud83d\udcbb",
    "description": "A pair-programming assistant for IDE / agent loops. Heavy on coding benchmarks, with a real-world agentic component (SWE-bench) and some weight on cost since coding loops burn tokens. A small slice goes to the recovery-rate reliability axis because real-world agent loops live or die on whether the model self-corrects.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "livecodebench",
        "weight": 0.3
      },
      {
        "benchmarkId": "swe-bench-verified",
        "weight": 0.25
      },
      {
        "benchmarkId": "aider-polyglot",
        "weight": 0.15
      },
      {
        "benchmarkId": "arena-elo",
        "weight": 0.15
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.1
      },
      {
        "benchmarkId": "recovery-rate",
        "weight": 0.05
      }
    ],
    "cost": {
      "weight": 0.2,
      "inputTokensPerRequest": 4000,
      "outputTokensPerRequest": 1500,
      "requestsPerDay": 200
    }
  },
  {
    "id": "local-coding",
    "name": "Local / Self-Hosted Coding",
    "emoji": "\ud83d\udda5\ufe0f",
    "description": "A coding assistant you run on your own hardware \u2014 no API bills, full privacy, works offline. Restricted to open-weight, self-hostable models and weighted toward practical coding and agentic ability, with a slice for the self-correction that smaller local models lean on inside a good harness. There is no cost axis because inference runs on hardware you already own.",
    "modality": "text",
    "selfHostableOnly": true,
    "weights": [
      {
        "benchmarkId": "livecodebench",
        "weight": 0.3
      },
      {
        "benchmarkId": "aider-polyglot",
        "weight": 0.25
      },
      {
        "benchmarkId": "swe-bench-verified",
        "weight": 0.2
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.1
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.1
      },
      {
        "benchmarkId": "recovery-rate",
        "weight": 0.05
      }
    ],
    "cost": null
  },
  {
    "id": "customer-support-bot",
    "name": "Customer Support Bot",
    "emoji": "\ud83d\udcac",
    "description": "A high-volume B2C chatbot. We weight Arena (human preference) and IFEval (does it follow your formatting instructions?) heavily, and lean on cost because volume is enormous. Reliability axes (format-adherence, safety-handling) are weighted modestly because a chatbot that answers off-format or false-refuses on benign requests is unusable regardless of how smart it is.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "arena-elo",
        "weight": 0.35
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.2
      },
      {
        "benchmarkId": "mmlu-pro",
        "weight": 0.1
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.1
      },
      {
        "benchmarkId": "format-adherence",
        "weight": 0.15
      },
      {
        "benchmarkId": "safety-handling",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.4,
      "inputTokensPerRequest": 1200,
      "outputTokensPerRequest": 250,
      "requestsPerDay": 5000
    }
  },
  {
    "id": "research-analyst",
    "name": "Research / Analyst",
    "emoji": "\ud83d\udd2c",
    "description": "A research assistant that reads long PDFs and answers nuanced questions. Weighted toward reasoning, knowledge, long-context, and a saturation-resistant frontier capability score so the ranking stays meaningful as MMLU-style evals saturate.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "gpqa-diamond",
        "weight": 0.2
      },
      {
        "benchmarkId": "humanitys-last-exam",
        "weight": 0.15
      },
      {
        "benchmarkId": "mmlu-pro",
        "weight": 0.1
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.25
      },
      {
        "benchmarkId": "simpleqa-verified",
        "weight": 0.1
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.2
      }
    ],
    "cost": {
      "weight": 0.1,
      "inputTokensPerRequest": 25000,
      "outputTokensPerRequest": 1500,
      "requestsPerDay": 50
    }
  },
  {
    "id": "math-tutor",
    "name": "Math / Reasoning Tutor",
    "emoji": "\ud83e\uddee",
    "description": "Tutors students through hard math. Reasoning-heavy; cost matters less because depth-of-thought is the value. Weighted toward contamination-resistant problem sets (FrontierMath, OTIS Mock AIME) since AIME 2024 and MATH-500 are now in most training corpora.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "frontiermath-t13",
        "weight": 0.3
      },
      {
        "benchmarkId": "otis-mock-aime",
        "weight": 0.2
      },
      {
        "benchmarkId": "aime-2024",
        "weight": 0.15
      },
      {
        "benchmarkId": "gpqa-diamond",
        "weight": 0.15
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.1
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.1,
      "inputTokensPerRequest": 1500,
      "outputTokensPerRequest": 3000,
      "requestsPerDay": 100
    }
  },
  {
    "id": "bulk-classification",
    "name": "Bulk Classification",
    "emoji": "\ud83c\udff7\ufe0f",
    "description": "Classifying / extracting fields from millions of documents. Quality barely matters past a threshold; cost dominates. Quality leans on broad everyday ability (LiveBench) and human preference, plus instruction and format following so outputs parse reliably.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "livebench-global",
        "weight": 0.3
      },
      {
        "benchmarkId": "arena-elo",
        "weight": 0.25
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.2
      },
      {
        "benchmarkId": "format-adherence",
        "weight": 0.15
      },
      {
        "benchmarkId": "mmlu-pro",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.65,
      "inputTokensPerRequest": 800,
      "outputTokensPerRequest": 80,
      "requestsPerDay": 50000
    }
  },
  {
    "id": "autonomous-agent",
    "name": "Autonomous Agent / Tool Use",
    "emoji": "\ud83e\udd16",
    "description": "An agent that runs multi-step jobs on its own: shell commands, code changes, tool calls. Weighted toward agentic benchmarks that grade finished tasks (Terminal-Bench, SWE-bench) plus novel-problem reasoning, with a slice for self-correction because a long run fails if the model cannot recover from its own mistakes. Cost matters because agent loops resend large contexts on every step.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "terminal-bench-2",
        "weight": 0.3
      },
      {
        "benchmarkId": "swe-bench-verified",
        "weight": 0.3
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.2
      },
      {
        "benchmarkId": "arc-agi-2",
        "weight": 0.1
      },
      {
        "benchmarkId": "recovery-rate",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.2,
      "inputTokensPerRequest": 6000,
      "outputTokensPerRequest": 1200,
      "requestsPerDay": 500
    }
  },
  {
    "id": "writing-content",
    "name": "Writing & Content",
    "emoji": "\u270d\ufe0f",
    "description": "Drafting articles, emails, product copy and summaries people will read. Human preference (Arena) leads because readers judge tone and clarity, with factual accuracy (SimpleQA) to limit made-up claims, broad everyday ability (LiveBench), and instruction and format following for briefs with length or style rules.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "arena-elo",
        "weight": 0.45
      },
      {
        "benchmarkId": "simpleqa-verified",
        "weight": 0.15
      },
      {
        "benchmarkId": "livebench-global",
        "weight": 0.15
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.15
      },
      {
        "benchmarkId": "format-adherence",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.25,
      "inputTokensPerRequest": 1500,
      "outputTokensPerRequest": 1200,
      "requestsPerDay": 1000
    }
  },
  {
    "id": "knowledge-assistant",
    "name": "Knowledge Q&A Assistant",
    "emoji": "\ud83d\udcda",
    "description": "An internal help desk or search assistant that answers questions from documents and general knowledge. Factual accuracy (SimpleQA) leads, with long-context retrieval for pasted documents, broad everyday ability, human preference for clear answers, and a slice of hard expert questions. Cost is weighted for steady daily volume.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "simpleqa-verified",
        "weight": 0.35
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.2
      },
      {
        "benchmarkId": "livebench-global",
        "weight": 0.2
      },
      {
        "benchmarkId": "arena-elo",
        "weight": 0.15
      },
      {
        "benchmarkId": "humanitys-last-exam",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.3,
      "inputTokensPerRequest": 3000,
      "outputTokensPerRequest": 500,
      "requestsPerDay": 2000
    }
  },
  {
    "id": "multimodal-document-qa",
    "name": "Multimodal Document Q&A",
    "emoji": "\ud83d\udcc4",
    "description": "Answering questions about scanned PDFs, charts, and screenshots. Relies on vision benchmarks plus reasoning.",
    "modality": "vision",
    "weights": [
      {
        "benchmarkId": "mmmu",
        "weight": 0.25
      },
      {
        "benchmarkId": "mmmu-pro",
        "weight": 0.2
      },
      {
        "benchmarkId": "mathvista",
        "weight": 0.25
      },
      {
        "benchmarkId": "gpqa-diamond",
        "weight": 0.2
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.2,
      "inputTokensPerRequest": 8000,
      "outputTokensPerRequest": 800,
      "requestsPerDay": 200
    }
  },
  {
    "id": "data-analyst",
    "name": "Data / BI Analyst",
    "emoji": "\ud83d\udcca",
    "description": "A spreadsheet / SQL / BI copilot that reads tables, answers analytical questions, and produces clean structured output. Weights a contamination-controlled data-analysis score heavily, plus long-context for large tables and instruction-following for rigid output formats.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "livebench-data-analysis",
        "weight": 0.45
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.25
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.2
      },
      {
        "benchmarkId": "gpqa-diamond",
        "weight": 0.1
      }
    ],
    "cost": {
      "weight": 0.25,
      "inputTokensPerRequest": 12000,
      "outputTokensPerRequest": 600,
      "requestsPerDay": 300
    }
  },
  {
    "id": "realtime-chat",
    "name": "Realtime Chat / Voice",
    "emoji": "\u26a1",
    "description": "A streaming consumer chat or voice assistant. Speed and time-to-first-token matter as much as raw quality \u2014 a slightly less smart model that responds instantly often beats a frontier model that pauses to think.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "tokens-per-second",
        "weight": 0.3
      },
      {
        "benchmarkId": "time-to-first-token",
        "weight": 0.25
      },
      {
        "benchmarkId": "arena-elo",
        "weight": 0.2
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.25
      }
    ],
    "cost": {
      "weight": 0.3,
      "inputTokensPerRequest": 800,
      "outputTokensPerRequest": 200,
      "requestsPerDay": 20000
    }
  },
  {
    "id": "marketing-image-gen",
    "name": "Marketing Image Generation",
    "emoji": "\ud83c\udfa8",
    "description": "Generating polished marketing visuals at scale. Weighs human preference and prompt fidelity against per-image cost.",
    "modality": "image-gen",
    "weights": [
      {
        "benchmarkId": "image-arena-elo",
        "weight": 0.55
      },
      {
        "benchmarkId": "image-prompt-adherence",
        "weight": 0.45
      }
    ],
    "cost": {
      "weight": 0.3,
      "inputTokensPerRequest": 0,
      "outputTokensPerRequest": 0,
      "requestsPerDay": 500
    }
  },
  {
    "id": "production-critical",
    "name": "Production Critical",
    "emoji": "\ud83d\udee1\ufe0f",
    "description": "Regulated or high-stakes drafting (legal contracts, healthcare notes, financial summaries). Reliability dominates: a model that occasionally outputs the wrong format or false-refuses on benign prompts is unusable here regardless of how smart it is. Quality is anchored to a saturation-resistant frontier capability score plus strict instruction-following. Cost weight is low because production buyers pay for trust.",
    "modality": "text",
    "weights": [
      {
        "benchmarkId": "format-adherence",
        "weight": 0.15
      },
      {
        "benchmarkId": "output-stability",
        "weight": 0.15
      },
      {
        "benchmarkId": "safety-handling",
        "weight": 0.15
      },
      {
        "benchmarkId": "recovery-rate",
        "weight": 0.1
      },
      {
        "benchmarkId": "ifeval",
        "weight": 0.15
      },
      {
        "benchmarkId": "frontier-composite",
        "weight": 0.25
      },
      {
        "benchmarkId": "long-context-ruler",
        "weight": 0.05
      }
    ],
    "cost": {
      "weight": 0.1,
      "inputTokensPerRequest": 6000,
      "outputTokensPerRequest": 1500,
      "requestsPerDay": 300
    }
  }
]
