{
  "post": {
    "title": "The real cost of GraphRAG: indexing, queries and keeping the graph fresh",
    "url": "https://bardak.ai/posts/graphrag-cost",
    "updated": "2026-10-29T18:55:35.884Z"
  },
  "citation": "The real cost of GraphRAG: indexing, queries and keeping the graph fresh, Ivan Bardak, bardak.ai, https://bardak.ai/posts/graphrag-cost",
  "datasets": [
    {
      "key": "tokens-used-to-index-one-56-000-token-book",
      "title": "Tokens used to index one 56,000-token book",
      "type": "chart",
      "source": "Xiang et al., When to use Graphs in RAG (GraphRAG-Bench), arXiv 2506.05690 v3, February 2026, Table 10",
      "caption": "Microsoft GraphRAG read the book's 56,000 tokens with about 12 times as many LLM tokens. The LazyGraphRAG row is a third-party implementation: Microsoft says its own version costs the same to index as vector RAG.",
      "csv": "https://bardak.ai/data/graphrag-cost/tokens-used-to-index-one-56-000-token-book.csv",
      "header": [
        "method",
        "Tokens (thousands)"
      ],
      "rows": [
        [
          "Microsoft GraphRAG",
          "655"
        ],
        [
          "LazyGraphRAG (third-party)",
          "591"
        ],
        [
          "KET-RAG",
          "518"
        ],
        [
          "LightRAG",
          "474"
        ],
        [
          "HippoRAG 2",
          "330"
        ],
        [
          "Fast-GraphRAG",
          "252"
        ],
        [
          "HippoRAG",
          "178"
        ],
        [
          "RAPTOR",
          "116"
        ],
        [
          "KGP",
          "89"
        ]
      ]
    },
    {
      "key": "illustrative-cost-to-index-with-microsoft-graphrag",
      "title": "Illustrative cost to index with Microsoft GraphRAG",
      "type": "benchmarkTable",
      "source": "Token counts: GraphRAG-Bench (2026); prices: OpenAI pricing page, accessed 26 October 2026",
      "caption": "My arithmetic: 535,832 input and 118,841 output tokens per 56,000-token book, scaled linearly. Batch processing halves these prices. Real corpora with more entities per page, or extra extraction passes, cost more.",
      "csv": "https://bardak.ai/data/graphrag-cost/illustrative-cost-to-index-with-microsoft-graphrag.csv",
      "header": [
        "Model",
        "Input $/1M",
        "Output $/1M",
        "Index 1M tokens ($)",
        "Index 100M tokens ($)"
      ],
      "rows": [
        [
          "GPT-4o mini",
          "0.15",
          "0.60",
          "2.71",
          "271"
        ],
        [
          "GPT-5 mini",
          "0.25",
          "2.00",
          "6.64",
          "664"
        ],
        [
          "GPT-4.1 mini",
          "0.40",
          "1.60",
          "7.22",
          "722"
        ],
        [
          "GPT-4 Turbo (original paper)",
          "10.00",
          "30.00",
          "159",
          "15,935"
        ]
      ]
    },
    {
      "key": "prompt-tokens-per-question-and-what-1-000-questions-cost",
      "title": "Prompt tokens per question, and what 1,000 questions cost",
      "type": "benchmarkTable",
      "source": "Xiang et al., GraphRAG-Bench (arXiv 2506.05690), Figures 8 and 9",
      "caption": "Novel dataset, GPT-4o mini as the answering model; cost is my arithmetic at $0.15 per million input tokens, excluding output. The paper's text gives lower figures for global search (up to about 40,000 tokens), but either way it is one to two orders of magnitude above vanilla RAG.",
      "csv": "https://bardak.ai/data/graphrag-cost/prompt-tokens-per-question-and-what-1-000-questions-cost.csv",
      "header": [
        "Method",
        "Prompt tokens per question",
        "$ per 1000 questions"
      ],
      "rows": [
        [
          "Vanilla RAG",
          "879",
          "0.13"
        ],
        [
          "HippoRAG 2",
          "1,008",
          "0.15"
        ],
        [
          "RAPTOR",
          "3,441",
          "0.52"
        ],
        [
          "Fast-GraphRAG",
          "4,204",
          "0.63"
        ],
        [
          "HippoRAG",
          "7,208",
          "1.08"
        ],
        [
          "Microsoft GraphRAG (local search)",
          "38,707",
          "5.81"
        ],
        [
          "LightRAG",
          "100,832",
          "15.12"
        ],
        [
          "Microsoft GraphRAG (global search)",
          "331,375",
          "49.71"
        ]
      ]
    },
    {
      "key": "graphrag-cost-model",
      "title": "GraphRAG cost model",
      "type": "benchmarkTable",
      "source": "",
      "caption": "Measure the factors on a sample of your own corpus and question log; benchmark numbers only give a starting range.",
      "csv": "https://bardak.ai/data/graphrag-cost/graphrag-cost-model.csv",
      "header": [
        "Cost",
        "How to estimate",
        "Typical driver"
      ],
      "rows": [
        [
          "Initial indexing",
          "Corpus tokens x LLM tokens per corpus token x price",
          "Model choice; extraction passes"
        ],
        [
          "Updates",
          "Changed tokens per month x the same factor, plus summary regeneration",
          "How often documents change"
        ],
        [
          "Full rebuilds",
          "Initial indexing cost x rebuilds per year",
          "Drift from incremental updates"
        ],
        [
          "Queries",
          "Questions per month x prompt tokens x price, per search mode",
          "Share of questions sent to global search"
        ],
        [
          "Embeddings and storage",
          "Vector index plus graph store",
          "Usually small"
        ],
        [
          "Evaluation",
          "Labelled question set, judging runs",
          "Repeated at every change"
        ],
        [
          "Engineering",
          "Pipeline, extraction tuning, monitoring",
          "Usually the largest line"
        ]
      ]
    },
    {
      "key": "appraising-the-evidence",
      "title": "Appraising the evidence",
      "type": "benchmarkTable",
      "source": "",
      "caption": "",
      "csv": "https://bardak.ai/data/graphrag-cost/appraising-the-evidence.csv",
      "header": [
        "Evidence",
        "Strength",
        "Main limitation"
      ],
      "rows": [
        [
          "GraphRAG-Bench token counts",
          "Independent, same corpus for every method",
          "One small book; a third-party LazyGraphRAG; text and figures disagree on global search"
        ],
        [
          "Edge et al.",
          "Primary source for GraphRAG",
          "Reports context tokens and time, not indexing tokens or dollars"
        ],
        [
          "Microsoft Research blogs",
          "The builders, with clear claims",
          "Own tests, models and cost units not always stated"
        ],
        [
          "LightRAG paper",
          "Clear argument on updates",
          "Estimates, not measurements, against a 2024 GraphRAG"
        ],
        [
          "My cost tables",
          "Transparent arithmetic",
          "Linear scaling from one book; list prices change"
        ]
      ]
    }
  ]
}