{
  "schema_version": "1.0",
  "retrieved_date": "2026-10-22",
  "sources": [
    {
      "id": "UP-S01",
      "title": "AI Index Report 2025, chapter 1: Research and Development",
      "publisher": "Stanford HAI AI Index",
      "published_date": "2025-04-07",
      "url": "https://hai.stanford.edu/assets/files/hai_ai-index-report-2025_chapter1_final.pdf",
      "locator": "Chapter PDF p.41, figure 1.3.22; equivalent full report p.64",
      "access_method": "Web PDF text opened and relevant page text read",
      "retrieved_date": "2026-10-22",
      "provenance": "Original report; analysis performed with Epoch AI and Artificial Analysis data",
      "limitations": [
        "Historical endpoints end in October 2024",
        "Duration in prose conflicts with reported endpoint dates",
        "Per-token API price, not end-to-end task cost"
      ]
    },
    {
      "id": "UP-S02",
      "title": "LLM inference prices have fallen rapidly but unequally across tasks",
      "authors": [
        "Ben Cottier",
        "Ben Snodin",
        "David Owen",
        "Tom Adamczewski"
      ],
      "publisher": "Epoch AI",
      "published_date": "2025-03-12",
      "url": "https://epoch.ai/data-insights/llm-inference-price-trends",
      "locator": "MMLU / threshold model GPT-3.5 table; Learn more about this graph / Data / Assumptions",
      "access_method": "Web primary page and methodology opened; primary GitHub CSV fetched through connector",
      "retrieved_date": "2026-10-22",
      "dataset_url": "https://github.com/epoch-research/llm-benchmark-efficiency/blob/2fdd9a8ca969d0b1b1068f09ab54de811f2e0ae2/results/default/lowest_price_models_above_previous_frontier/lowest_price_models_data.csv",
      "dataset_commit": "2fdd9a8ca969d0b1b1068f09ab54de811f2e0ae2",
      "dataset_blob_sha": "c97bc4c9e0e10d07a4003cd8a3d6b64c3b7df1c1",
      "dataset_file_commit_date": "2025-03-13T18:21:52Z",
      "license": "Epoch website states Creative Commons Attribution; only factual selected rows retained",
      "limitations": [
        "Benchmark coverage is incomplete",
        "Benchmark prompting and contamination can affect equivalence",
        "Reasoning models excluded",
        "Price mix fixed at 3 input : 1 output",
        "First-party prices preferred; otherwise provider median; shortest context where applicable"
      ]
    },
    {
      "id": "UP-S03",
      "title": "AI Index Report 2026",
      "publisher": "Stanford HAI AI Index",
      "published_date": "2026-04",
      "url": "https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf",
      "locator": "Section 1.4, printed p.36, Inference; full text search for price and inference terms",
      "access_method": "Web PDF size limit encountered; public PDF downloaded to temporary storage with curl and text-extracted with pdftotext; full extracted text searched and relevant section read",
      "retrieved_date": "2026-10-22",
      "sha256": "fff5ac1df971cee7daf847fa5ed00dd7444db693092821c3e691debc1eaada09",
      "limitations": [
        "No updated fixed-capability token-price series found in this report check",
        "Energy estimates are not API price measurements",
        "No claim of exhaustive absence across all 2026 research"
      ]
    },
    {
      "id": "UP-S04",
      "title": "Learning to reason with LLMs",
      "publisher": "OpenAI",
      "published_date": "2024-09-12",
      "url": "https://openai.com/index/learning-to-reason-with-llms/",
      "locator": "Evals AIME paragraph and Appendix A AIME (2024) o1 column",
      "access_method": "Official primary article and appendix table opened",
      "retrieved_date": "2026-10-22",
      "limitations": [
        "Vendor-reported proprietary research-model evaluation",
        "Historical configuration; not o1-preview or current product pricing",
        "No dollar or FLOPs budget per configuration disclosed",
        "Sample count and selection method change together",
        "Average per 15-question exam; exact repetitions and uncertainty for this contrast not supplied in article"
      ]
    },
    {
      "id": "UP-S05",
      "title": "Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters",
      "authors": [
        "Charlie Snell",
        "Jaehoon Lee",
        "Kelvin Xu",
        "Aviral Kumar"
      ],
      "publisher": "arXiv / authors at UC Berkeley and Google DeepMind",
      "published_date": "2024-08-06",
      "version": "arXiv:2408.03314v1",
      "url": "https://arxiv.org/html/2408.03314v1",
      "locator": "Sections 3.2, 4, 5.3, 6.2; figures 4 and 8; Appendix C",
      "access_method": "Primary paper HTML read",
      "retrieved_date": "2026-10-22",
      "limitations": [
        "Fine-tuned PaLM 2-S* and MATH 500 test set",
        "Difficulty-estimation sampling overhead excluded",
        "Generation budgets / compute proxies, not measured API-dollar savings",
        "Limited progress on hardest questions; some over-optimisation harms easy questions",
        "No exact numerical accuracy endpoints transcribed from curves"
      ]
    }
  ]
}