2026-10-22-sources
{
"schema_version": 1,
"accessed": "2026-10-22",
"scope": "Cost economics research; primary sources",
"sources": [
{
"id": "C01",
"title": "OpenAI API pricing",
"publisher": "OpenAI",
"url": "https://developers.openai.com/api/docs/pricing",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official live rate card",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"GPT-6.1 Sol Standard short-context prices: input 2, output 10, cached reads 0.10, 30-minute writes 2.50 and 12-hour writes 5 USD per million tokens. Hosted search and storage have separate charges."
],
"limitations": [
"List rates only; context and processing selectors matter; tax, private contracts and regional premiums not included. Search-cache results differed from opened page; opened page used."
],
"local_card": "2026-10-22-C01.md"
},
{
"id": "C02",
"title": "Claude Platform pricing",
"publisher": "Anthropic",
"url": "https://platform.claude.com/docs/en/about-claude/pricing",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official live rate card",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Sonnet 5.5 input/output rates are 2/10 USD per million tokens. Haiku 5.5 uses 0.10/0.50 up to 100,000 prompt tokens and 0.50/2.50 above. Batch has separate lower rates."
],
"limitations": [
"First-party global default; partner and regional rates differ. Tokens per text differ by tokenizer. No quality equivalence implied."
],
"local_card": "2026-10-22-C02.md"
},
{
"id": "C03",
"title": "Gemini Developer API pricing",
"publisher": "Google",
"url": "https://ai.google.dev/gemini-api/docs/pricing",
"published": null,
"updated": "2026-10-19",
"accessed": "2026-10-22",
"source_type": "Official live rate card",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Argon Standard current promotional input/output rates are 2/10 USD per million tokens through November 30, 2026; page announces later rates. Flash-Lite 3.5 has Standard 0.30/2.50 and Batch 0.15/1.25."
],
"limitations": [
"Developer API paid tier only; not a Vertex/enterprise quote. Tool and explicit-cache storage costs are additional; preview/promotional terms can change."
],
"local_card": "2026-10-22-C03.md"
},
{
"id": "C04",
"title": "Prompt caching",
"publisher": "OpenAI",
"url": "https://developers.openai.com/api/docs/guides/prompt-caching",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official implementation documentation",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Caching reuses an unchanged prefix. GPT-5.6+ has a 1,024-visible-token minimum and separately priced writes. Verify actual read/write counters."
],
"limitations": [
"Requires reuse while cache is valid; cache read discount is not a whole-application saving; model and retention settings differ."
],
"local_card": "2026-10-22-C04.md"
},
{
"id": "C05",
"title": "Batch API",
"publisher": "OpenAI",
"url": "https://developers.openai.com/api/docs/guides/batch",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official implementation documentation",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Asynchronous batch processing has 50% lower model costs versus synchronous API with a 24-hour processing window. Expired jobs are explicitly possible."
],
"limitations": [
"Not an interactive SLA or guarantee every request succeeds. Eligibility, result reconciliation and failed/expired items require handling."
],
"local_card": "2026-10-22-C05.md"
},
{
"id": "C06",
"title": "Reasoning models",
"publisher": "OpenAI",
"url": "https://developers.openai.com/api/docs/guides/reasoning",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official implementation documentation",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Reasoning tokens are billed within output tokens. Output limits include reasoning and can be reached before a visible answer."
],
"limitations": [
"Do not add reasoning tokens twice to an output total. Behavior and context handling vary by model."
],
"local_card": "2026-10-22-C06.md"
},
{
"id": "C07",
"title": "Context caching",
"publisher": "Google",
"url": "https://ai.google.dev/gemini-api/docs/generate-content/caching",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official implementation documentation",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Implicit cache hits depend on prefix reuse; explicit cache cost also includes token storage time. Explicit caching is unsupported for Argon."
],
"limitations": [
"Legacy GenerateContent documentation; use actual model/endpoint support. No universal hit-rate guarantee."
],
"local_card": "2026-10-22-C07.md"
},
{
"id": "C08",
"title": "NVIDIA NIM LLM benchmarking overview",
"publisher": "NVIDIA",
"url": "https://docs.nvidia.com/nim/benchmarking/llm/overview",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official benchmark methodology",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Inference includes queue, prefill and decode. Input/output lengths and concurrency matter; benchmark cost and latency alongside separately validated accuracy."
],
"limitations": [
"Vendor methodology; not proof a model meets Knauf quality or production demand."
],
"local_card": "2026-10-22-C08.md"
},
{
"id": "C09",
"title": "Lambda on-demand GPU instances",
"publisher": "Lambda",
"url": "https://lambda.ai/instances",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official live infrastructure rate card",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"One-GPU H100 PCIe 80 GB plan lists 3.29 USD/GPU/hour plus tax; one-GPU H100 SXM lists 4.29."
],
"limitations": [
"Specific SKU and one-GPU plan; geography not specified in rate table; capacity availability, enterprise SLA and total operations cost not established."
],
"local_card": "2026-10-22-C09.md"
},
{
"id": "C10",
"title": "How to Build a Generative AI Cost and Usage Tracker",
"publisher": "FinOps Foundation",
"url": "https://www.finops.org/wg/how-to-build-a-generative-ai-cost-and-usage-tracker/",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Practitioner foundation guidance",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Recommends request attribution and date-aware model pricing records; token-based inference tracking is only one part of the cost stack."
],
"limitations": [
"Its example prices are not used here. Example SQL is illustrative, not production-ready billing code."
],
"local_card": "2026-10-22-C10.md"
},
{
"id": "C11",
"title": "Optimize resource consumption to minimize hosting costs",
"publisher": "Amazon Web Services",
"url": "https://docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens/gencost02-bp02.html",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official architecture guidance",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Profile demand before right-sizing hosting to performance requirements; predictable workloads can support scheduled shutdown and capacity commitments."
],
"limitations": [
"Cloud vendor guidance. Any commitment must be evaluated against measured utilization and operational requirements."
],
"local_card": "2026-10-22-C11.md"
},
{
"id": "C12",
"title": "Architecting generative AI applications for production",
"publisher": "Amazon Web Services",
"url": "https://docs.aws.amazon.com/prescriptive-guidance/latest/gen-ai-lifecycle-operational-excellence/preprod-architecting.html",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official architecture guidance",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Production cost models should include traffic patterns, input/output tokens, model prices, infrastructure, vector storage and guardrails."
],
"limitations": [
"No numeric savings assumed; architectures and routing must be validated on the actual task."
],
"local_card": "2026-10-22-C12.md"
},
{
"id": "C13",
"title": "OpenAI API changelog: GPT-6.1 Sol release",
"publisher": "OpenAI",
"url": "https://developers.openai.com/api/docs/changelog",
"published": "2026-09-29",
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official dated release entry",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"September 29 entry states GPT-6.1 Sol Standard prices apply through 272,000 input tokens; September 22 entry states the same short-context boundary for Luna."
],
"limitations": [
"Dated release entry establishes boundary; current pricing page remains the rate authority."
],
"local_card": "2026-10-22-C13.md"
},
{
"id": "C14",
"title": "MLPerf Inference Datacenter benchmark methodology",
"publisher": "MLCommons",
"url": "https://mlcommons.org/benchmarks/inference-datacenter/",
"published": null,
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official benchmark suite methodology",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"Benchmarks define datasets, quality targets, serving scenarios and performance metrics."
],
"limitations": [
"A standardized benchmark result does not establish Knauf task quality, peak-load SLA or total cost."
],
"local_card": "2026-10-22-C14.md"
},
{
"id": "C15",
"title": "Introducing the MLPerf End-to-End RAG Inference Benchmark",
"publisher": "MLCommons",
"url": "https://mlcommons.org/2026/08/endtoend-inference/",
"published": "2026-08-26",
"updated": null,
"accessed": "2026-10-22",
"source_type": "Official benchmark announcement",
"evidence_status": "Opened official primary page and inspected relevant content",
"key_paraphrased_findings": [
"RAG benchmark measures ingestion in documents/second and QnA in tasks/second; its first version is Offline, with all requests available together."
],
"limitations": [
"Offline throughput is not online user experience or tail-latency evidence."
],
"local_card": "2026-10-22-C15.md"
}
]
}