# The Cache Wars 2

Open cache-wars-2.pdf or cache-wars-2.html. The same academic manuscript contains the original 11-section structure, seven original chart types and twelve numbered tables, plus six Codex charts (C1–C6) and eight Codex tables (C1–C8).

Both Claude Sonnet 5 and GPT-6 Astra have per-request results, pause accounting, price lists, common-content cost analysis, one-hour and multi-hour projections, monthly and annual costs, team scale, cache-loss exposure, and subscription/API accounting. AGNT leads the observed post-pause cache-read percentage on both tracks (90.45% / 88.42%). Equal content with equal prefix availability costs the same; the percentage result is not a universal price ranking.

Original study: https://agnt.gg/whitepapers/the-cache-wars-prompt-cache-efficiency-llm-agent-harnesses

Verify: `python artifacts/scripts/verify.py`

Rebuild: `python artifacts/scripts/build.py --output NEW_DIRECTORY`

Compile the emitted LaTeX twice with pdfLaTeX. Matplotlib 3.10.3 generates the vector figures. The Codex model includes the current 272,000-input-token long-context threshold. Its loss schedule is a stated sensitivity scenario, not an inferred miss probability or a subscription bill.

SHA256SUMS.txt covers every payload except itself. The manifest is an integrity reference, not a signature or trusted timestamp. LIVE-REPRODUCTION.md contains credential requirements and pinned invocation settings. Private account context and credentials are excluded.