{ "schemaVersion": 1, "retrievedOn": "2026-10-22", "asOfDate": "2026-10-22", "purpose": "Scoped public-primary-source claim audit for Knauf AI cost optimisation; no experimental reproduction", "quotesStored": false, "claims": [ { "id": "AG01", "placement": "main-candidate", "slideClaim": "Agentic architectures can multiply token demand.", "approvedCaption": "Anthropic observed ~4× tokens for agents and ~15× for multi-agent systems versus chats in its own data.", "metric": "total tokens relative to chat interactions", "chartData": [ { "label": "Chat", "value": 1 }, { "label": "Agent", "value": 4, "approximate": true }, { "label": "Multi-agent", "value": 15, "approximate": true } ], "source": { "title": "How we built our multi-agent research system", "publisher": "Anthropic", "publishedOn": "2025-06-13", "url": "https://www.anthropic.com/research/multi-agent-research-system", "locator": "Benefits of a multi-agent system, paragraphs 5–7", "type": "first-party engineering observation" }, "supports": "Architecture and task decomposition materially change resource consumption.", "doesNotSupport": "15× against single agents; universal or future bill multipliers; matched-task causal comparison.", "methodLimitations": "Population, sample size and precise token-accounting protocol are undisclosed.", "counterevidence": "The same post reports 90.2% improvement on an internal research evaluation; extra effort can create valuable capability. Do not infer a joint cost/quality experiment.", "verdict": "Supported as an attributed, approximate observation only" }, { "id": "AG02", "placement": "main-candidate", "slideClaim": "More reasoning does not reliably buy a better outcome.", "approvedCaption": "Higher reasoning effort produced equal or lower accuracy in 21 of 36 matched HAL comparisons.", "metric": "model–scaffold–benchmark comparisons with no accuracy improvement", "chartData": { "noImprovement": 21, "improvement": 15, "total": 36 }, "source": { "title": "Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation", "publisher": "Princeton-led research team", "publishedOn": "2025-10-13", "version": "arXiv:2510.11977v1", "url": "https://arxiv.org/html/2510.11977v1", "locator": "Figure 3; §4.1 item 5; Appendix A4", "type": "independent benchmark study" }, "method": "Four model pairs across fixed scaffolds/benchmarks: Claude 3.7, Sonnet 4, Opus 4.1 no/high reasoning; o4-mini low/high. Broader study: 21,730 rollouts, 9 benchmarks, 9 models.", "doesNotSupport": "58% of tasks or spending wasted; causal effect of reasoning on arbitrary enterprise workloads.", "counterevidence": "Accuracy and token use correlate positively on 6/9 benchmarks. Reasoning can help.", "methodLimitations": "Incomplete evaluation matrix; small SWE-bench subset; provider settings differ. SWE-Agent cost accounting omits cache discounts, limiting its dollar comparisons; no reliable latency claim.", "verdict": "Supported under the tested model–scaffold–benchmark settings" }, { "id": "AG03", "placement": "backup-candidate", "approvedCaption": "Context editing reduced token consumption by 84% in Anthropic's 100-turn web-search evaluation.", "metric": "token consumption relative to no context editing", "chartData": [ { "label": "Without editing", "value": 100 }, { "label": "With editing", "value": 16 } ], "source": { "title": "Managing context on the Claude Developer Platform", "publisher": "Anthropic", "publishedOn": "2025-09-29", "url": "https://claude.com/resources/articles/context-management", "locator": "Performance improvements with context management", "type": "first-party internal evaluation" }, "supports": "Workflow/context engineering can reduce consumption while enabling completion of long workflows.", "doesNotSupport": "84% savings on arbitrary workloads or an 84% reduction in total bills.", "methodLimitations": "Vendor internal web-search evaluation; sample count, detailed distribution and full protocol undisclosed. Completion compared with workflows exhausting context, not a quality-equivalent financial comparison.", "verdict": "Supported as a scoped vendor example" }, { "id": "AG04", "placement": "backup-candidate", "approvedCaption": "Repeated coding-agent runs on the same issue showed substantial cost variability.", "metric": "max/min token-cost ratio across four independent runs, averaged across issues", "chartData": { "approximateMeanMaxMinRatio": 2, "exactPerModelNumbersAvailable": false }, "source": { "title": "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks", "publisher": "Bai et al.; Stanford/Michigan and collaborators", "publishedOn": "2026-04-24", "revisedOn": "2026-10-02", "version": "arXiv:2604.22750v3", "url": "https://arxiv.org/html/2604.22750v3", "locator": "§2; §3 Figure2; §7 limitations", "type": "working paper" }, "method": "OpenHands, 500 SWE-bench Verified issues, eight models, four runs/issue/model; full history carried forward unchanged.", "supports": "Budget distributions and runtime limits matter even with fixed tasks/models.", "doesNotSupport": "30× is typical; causal harm from extra tokens; equivalent-workload 1000× demand comparison.", "methodLimitations": "One scaffold/domain; association rather than randomized effort intervention; recent preprint. The 30× headline is an extreme; approximately 2× is the repeated-run summary.", "verdict": "Supported within this coding-agent setup; use approximate variability or verbal explanation" }, { "id": "AG05", "placement": "backup-candidate", "approvedCaption": "Cost-aware prompt optimisation lowered variable inference cost while preserving benchmark accuracy.", "metric": "variable inference cost versus default DSPy implementations", "chartData": [ { "model": "GPT-3.5", "reductionPercent": 53 }, { "model": "Llama-3-70B", "reductionPercent": 41 } ], "source": { "title": "AI Agents That Matter", "publisher": "Kapoor et al., Princeton University", "publishedOn": "2024-07-01", "version": "arXiv:2407.01502v1", "url": "https://arxiv.org/html/2407.01502v1", "locator": "§3.1–3.2; Figure2; Appendix B", "type": "independent benchmark study; journal publication reported 2025" }, "method": "HotPotQA multi-hop retrieval, 100 optimisation examples, 200 evaluation examples; five repetitions. Optimise temperature and few-shot selection/count/format.", "doesNotSupport": "Same savings on current models or Knauf; end-to-end business cost reduction.", "counterevidence": "One-time optimisation adds fixed cost. Study reports total-cost break-even around 1,350 tasks, specific to its setup.", "verdict": "Supported as a historical proof of mechanism, not a current savings forecast" }, { "id": "AG06", "placement": "backup-alternative", "approvedCaption": "Anthropic's 2026 guidance reports multi-agent implementations typically using 3–10× tokens versus single-agent approaches on equivalent tasks in its testing.", "metric": "tokens versus single-agent approach", "source": { "title": "Building multi-agent systems: When and how to use them", "publisher": "Anthropic", "publishedOn": "2026-01-23", "url": "https://claude.com/resources/articles/building-multi-agent-systems-when-and-how-to-use-them", "locator": "The case for starting with a single agent", "type": "first-party engineering guidance" }, "supports": "Duplicated context, coordination and handoffs can create material overhead.", "doesNotSupport": "Same baseline as AG01 or proof of its multipliers; universal model bills.", "methodLimitations": "No disclosed sample size, workload mix, token-accounting protocol or task-level table.", "counterevidence": "The article identifies context isolation, parallelism and specialisation as cases where extra agents can justify their overhead.", "verdict": "Supported as attributed vendor guidance; avoid displaying simultaneously with AG01" } ], "heldBackSources": [ { "title": "The Price Reversal Phenomenon: When Cheaper Reasoning Models Cost More", "url": "https://arxiv.org/html/2603.23971v2", "version": "2026-05-28 v2", "finding": "106/336 model-pair/dataset comparisons reverse input-plus-output list-price ranking and observed token-charge ranking.", "reasonHeldBack": "Explicitly excludes quality; composite list-price proxy is workload-neutral; scope is 12 datasets, not enterprise procurement. Source prose says one in five despite 32%. Abstract 28× maximum should not be promoted without reconciling body examples. v1 has different models/tasks and 21.8%; never mix versions.", "use": "Mechanism corroboration in research notes only" }, { "title": "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains", "url": "https://arxiv.org/abs/2406.12045", "reasonHeldBack": "Original 2024 model results are unsuitable as current 2026 capability estimates. pass^k measures all-k reliability, not at-least-one-success retry pass@k." } ], "accountingRule": "Token counts are not bills. Reprice uncached input, cache writes, cache reads, output/reasoning separately; add paid tools, hosting and review as appropriate. Compare cost per accepted outcome at fixed quality/latency constraints.", "generalisationRule": "These sources establish mechanisms and workload-specific observations, not the future level of Knauf spend or a guaranteed savings percentage.", "humanReviewFinding": "No transferable numerical review-cost estimate established in this audit; measure actual reviewer minutes and acceptance/rework locally." }