{
  "models": {
    "Llama 3.1 8B": {
      "organization": "Meta",
      "host": "AWS",
      "host_region": {
        "region_name": "AWS US-East-1 (Virginia)",
        "grid_carbon_g_per_kwh": 327,
        "water_stress_score": 2.56
      },
      "speedTier": "fast",
      "tokensPerSecond": 150,
      "short": {
        "energy_wh": 0.0517,
        "water_ml": 0.2727,
        "carbon_g": 0.0155
      },
      "medium": {
        "energy_wh": 0.1723,
        "water_ml": 0.9096,
        "carbon_g": 0.0517
      },
      "long": {
        "energy_wh": 0.443,
        "water_ml": 2.3383,
        "carbon_g": 0.1329
      },
      "insight": "Open-weight 8B model — roughly 1/30th the energy of frontier reasoning models for the same prompt."
    },
    "GPT-4o (Aug)": {
      "organization": "OpenAI",
      "host": "Azure",
      "host_region": {
        "region_name": "Azure East US (Virginia)",
        "grid_carbon_g_per_kwh": 327,
        "water_stress_score": 2.56
      },
      "speedTier": "medium",
      "tokensPerSecond": 80,
      "short": {
        "energy_wh": 0.4187,
        "water_ml": 1.9333,
        "carbon_g": 0.1423
      },
      "medium": {
        "energy_wh": 1.3541,
        "water_ml": 6.2532,
        "carbon_g": 0.4604
      },
      "long": {
        "energy_wh": 2.2215,
        "water_ml": 10.2584,
        "carbon_g": 0.7553
      },
      "insight": "OpenAI's August 2024 mid-tier production model — typical per-query cost for everyday consumer prompts."
    },
    "Claude 3.7 Sonnet": {
      "organization": "Anthropic",
      "host": "Anthropic",
      "host_region": {
        "region_name": "AWS US-East-1 (Virginia)",
        "grid_carbon_g_per_kwh": 327,
        "water_stress_score": 2.56
      },
      "speedTier": "medium",
      "tokensPerSecond": 60,
      "short": {
        "energy_wh": 0.9503,
        "water_ml": 5.0155,
        "carbon_g": 0.2851
      },
      "medium": {
        "energy_wh": 2.9886,
        "water_ml": 15.7736,
        "carbon_g": 0.8966
      },
      "long": {
        "energy_wh": 5.6714,
        "water_ml": 29.9329,
        "carbon_g": 1.7014
      },
      "insight": "Anthropic's mid-tier production model — comparable energy profile to GPT-4o but slightly higher water and carbon footprint."
    },
    "GPT-5 (minimal)": {
      "organization": "OpenAI",
      "host": "Azure",
      "host_region": {
        "region_name": "Azure East US (Virginia)",
        "grid_carbon_g_per_kwh": 327,
        "water_stress_score": 2.56
      },
      "speedTier": "medium",
      "tokensPerSecond": 70,
      "short": {
        "energy_wh": 0.7894,
        "water_ml": 3.6451,
        "carbon_g": 0.2684
      },
      "medium": {
        "energy_wh": 2.1027,
        "water_ml": 9.7101,
        "carbon_g": 0.7149
      },
      "long": {
        "energy_wh": 6.3831,
        "water_ml": 29.4764,
        "carbon_g": 2.1703
      },
      "insight": "GPT-5 with the lowest reasoning effort tier — most efficient setting on the same model weights."
    },
    "GPT-5 (high)": {
      "organization": "OpenAI",
      "host": "Azure",
      "host_region": {
        "region_name": "Azure East US (Virginia)",
        "grid_carbon_g_per_kwh": 327,
        "water_stress_score": 2.56
      },
      "speedTier": "slow",
      "tokensPerSecond": 40,
      "short": {
        "energy_wh": 13.6825,
        "water_ml": 63.1837,
        "carbon_g": 4.652
      },
      "medium": {
        "energy_wh": 16.4516,
        "water_ml": 75.9709,
        "carbon_g": 5.5935
      },
      "long": {
        "energy_wh": 25.5958,
        "water_ml": 118.1977,
        "carbon_g": 8.7026
      },
      "insight": "GPT-5's high reasoning tier — about 6× more energy than the same model on minimal effort, almost entirely from extended chain-of-thought generation."
    },
    "DeepSeek R1 (DeepSeek) [128k]": {
      "organization": "DeepSeek",
      "host": "DeepSeek",
      "host_region": null,
      "speedTier": "slow",
      "tokensPerSecond": 30,
      "short": {
        "energy_wh": 19.2507,
        "water_ml": 134.0018,
        "carbon_g": 11.5504
      },
      "medium": {
        "energy_wh": 24.5958,
        "water_ml": 171.2082,
        "carbon_g": 14.7575
      },
      "long": {
        "energy_wh": 29.0782,
        "water_ml": 202.4102,
        "carbon_g": 17.4469
      },
      "insight": "Vendor-hosted DeepSeek R1 — about 4× less energy than the same model hosted on Azure infrastructure; hosting choice matters as much as model choice for inference footprint."
    }
  },
  "tiers": {
    "short": {
      "tokens": 300,
      "label": "Short prompt",
      "prompt_example": "What's 2+2?",
      "response_example": "4."
    },
    "medium": {
      "tokens": 1000,
      "label": "Medium prompt",
      "prompt_example": "Explain the difference between SQL and NoSQL databases.",
      "response_example": "SQL databases store data in structured tables with strict schemas — examples include MySQL and PostgreSQL. They excel at complex relational queries and transactions. NoSQL databases like MongoDB use flexible document or key-value stores, which scale horizontally and handle unstructured data well."
    },
    "long": {
      "tokens": 1500,
      "label": "Long reasoning prompt",
      "prompt_example": "What's the most efficient way to write a recursive function that calculates the Fibonacci sequence?",
      "response_example": "Naive recursion for Fibonacci has O(2ⁿ) complexity due to repeated subproblems. The most efficient approach uses memoization:\n\nfunction fib(n, memo = {}) {\n  if (n < 2) return n;\n  if (memo[n]) return memo[n];\n  return memo[n] = fib(n-1, memo) + fib(n-2, memo);\n}\n\nThis reduces time and space complexity to O(n)."
    }
  },
  "equivalences": {
    "energy_wh_per_unit": {
      "value": 1,
      "comparison": "30 seconds of phone charging"
    },
    "water_ml_per_unit": {
      "value": 1,
      "comparison": "0.2 teaspoons of drinking water"
    },
    "carbon_g_per_unit": {
      "value": 1,
      "comparison": "7 metres of average car driving"
    }
  },
  "defaults": {
    "model": "GPT-4o (Aug)",
    "tier": "long"
  }
}