May 2026 token telemetry was cache-dominant: 82.9% of 73,950,305 recorded tokens were cache reads, suggesting persistent workflows shift the economic unit from cost per token to cost per completed artifact¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q1
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q1 i? In the strict May 1–May 25 trajectory subset (627 model-completed events), cache-read tokens represented 61,278,669 of 73,950,305 recorded tokens, a cache-dominance ratio of 0.829. → Anas H. Alzahrani 2026 q1 i? The author interprets the cache-heavy profile as evidence that persisted context was central to the workflow and that per-token accounting alone is a poor proxy for value. → Anas H. Alzahrani 2026 (2)

Evidence¶

Anas H. Alzahrani 2026¶

Anas H. Alzahrani, MD, PhD, MPH. (2026). Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study. arXiv preprint. https://arxiv.org/abs/2605.26870 doi:10.48550/arxiv.2605.26870

q1 · i? · design · r2

Token-telemetry analysis of the strict May 1–May 25 model-completed trajectory subset (627 events) in the single-investigator case study. The article states token telemetry is usage telemetry, not invoice cost, and covers only a partial month.

"Cache-read tokens represented 61,278,669 of the total, yielding a cache-dominance ratio of 0.829 (CDR = 82.9%), compared with 10,697,394 input tokens, 754,633 output tokens, and 1,219,609 cache-write tokens."

Anas H. Alzahrani 2026 (2)¶

Anas H. Alzahrani, MD, PhD, MPH. (2026). Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study. arXiv preprint. https://arxiv.org/abs/2605.26870

q1 · i? · design · r2

Authors' interpretation (type e) of the cache-dominance finding in the Resource Profile section, attributed to reused context rather than isolated fresh inference. Provider routes included OpenAI Codex-route, OpenAI-route, and Anthropic-route usage.

"The cache-heavy profile is an empirical signal that persisted context was central to the workflow. It also shows why per-token accounting alone is a poor proxy for value."

Discussion¶

Related Claims¶

• Within May completions, cache-read tokens per completion were positively associated with output tokens per completion (Pearson r = 0.47 on log-transformed counts; Spearman rho = 0.70) — related
• Human-written AGENTS.md context files improve agent performance by approximately 4 per cent, while LLM-generated context files decrease performance by 3 per cent and increase token cost by over 20 per cent, as the article reports citing recent research — related
• The memory layer recorded 889 failure, verification, correction, or protocol-proxy events (9.26 per active day), making governance part of the operating environment — related
• A persistent agentic research environment sustained near-daily use over 115 days, with 75,671 de-duplicated telemetry records across 96 active days (active-day fraction 0.835) — related
• Purpose-trained single-pass generation completes a slide in a median of 17 seconds and an interactive page in 59 seconds over 220k production requests, and delivers artifacts at 15–22× lower per-artifact API cost than flagship models
• Accuracy-cost value frontier: at $0.10 per million tokens, CDPK performance rose from 50% (April 2024) to 82% (June 2025)