Skip to content
MarginaliaDocsGitHub
Contents
handovers/2026-08-18-session-c11-in-numbers.md

Rendered from the repository — the file stays the source of truth.

Session C11 — Marginalia in numbers (2026-08-18)

Goal

Roadmap lane C, session C11 (docs/in-numbers): a public product/in-numbers.md answering two questions with dated, source-cited numbers — (1) what did AI-agent development of this prototype cost in API-equivalent inference spend, per session/lane and total, computed from the session transcripts’ usage records; (2) what does operating it cost today, and what would 10 users cost per month, derived from the code’s actual constants at freshly fetched provider prices. Docs-only; no product code changed.

What was done

  • product/in-numbers.md (NEW). Development half: per-session table (28 transcripts), lane subtotals, grand total ≈ $854.12 API-equivalent / ~622M processed tokens, all on claude-fable-5, priced at the published rates fetched 2026-08-18 ($10 in / $12.50 5m-write / $20 1h-write / $1 read / $50 out per MTok). Ops half: the current fixed bill (€4.99/mo — Edge Services Starter + 1 extra pipeline; everything else Free-tier/scale-to-zero at ~€0), the demo-mode derivation (min_scale=1 ≈ €34.8/mo from fetched serverless prices × the container’s TF sizing, matching the recorded €35), and a two-scenario 10-user model: (a) quota ceiling ≈ €350 + $250/mo (the A6 quotas’ hard upper bound), (b) a labeled realistic assumption (10 msgs/user/day, 2 sources/week, 2 overviews/user/mo) ≈ €8.4/mo — inference (€3.4) is smaller than the Edge Services subscription. Non-inference ceilings noted: Supabase Free 1 GB storage binds first (SEC-10; one user’s caps allow 20 GB), Pro $25/mo as the forced step, Azure F0’s 0.5M chars/mo caps audio 30× below app quotas, closed-signup caveat stated. Honesty rails throughout: every figure is computed / fetched (URL + 2026-08-18 date + “prices change”) / labeled assumption; aggregates only, no transcript content anywhere.
  • Docs app registration (the only apps/docs touches): src/pages/product/in-numbers.astro (same DocPage pattern as scope.astro), one nav item in the Product group (src/nav.ts), and one fixed rule in src/canonical-links.mjs mapping in-numbers.md → /product/in-numbers/ so future canonical links resolve (C12 plans an NF-15 → in-numbers pointer).
  • Correct-the-record: dated correction blockquote appended to handovers/2026-08-18-foreman-handover.md — its informal “~$300 through day 2” is superseded by the computed ≈ $854 (≈ $700 at that estimate’s own cutoff); its “foreman ≈ 40%” share held (40.1% computed).

Method (so the page can be refreshed)

Reproduced in the page’s appendix; operationally: throwaway scripts in the session scratchpad (NOT committed) globbed ~/.claude/projects/-home-donheidi-code-notebooklm-clone/*.jsonl, kept type == "assistant" lines with message.usage, deduplicated by message.id (usage repeats on one line per content block), skipped <synthetic>, summed uncached input / cache_creation.ephemeral_{5m,1h} / cache-read / output tokens per model id per file, and priced at the fetched rates (5m write = 1.25× input, 1h = 2×, read = 0.1×). Labels came from executing session <ID> in each file’s first 200 KB; the two unlabeled files are the foreman sessions (identified by timestamp spans matching the foreman handovers — file boundaries at exactly 18:00 day 2). server_tool_use.web_search_requests summed to zero, so no search charges. Ops prices: Anthropic pricing page, Scaleway model-as-a-service / serverless / network pages, Supabase pricing, Azure pricing page (F0 allowance) and the Azure Retail Prices API for the S1 neural-TTS $/char in swedencentral (the public pricing page renders prices via JS and shows “$-” to fetchers — the Retail Prices API is the reliable published source).

Findings worth keeping

  • One transcript carries both B3 and B4 — those roadmap sessions ran as a single physical session (the B4 brief appears mid-file); reported as one row. Check whole files, not just heads, when a session seems missing.
  • The two recorded false starts (B1, D2) cost $3.77 combined.
  • Cache reads are 72% of dev spend; uncached the same work ≈ $6,700.
  • The C11 row measures itself mid-session (it was ~$17 at computation time) and undercounts; the total is a floor. Re-running the appendix method after this merge refreshes it.

Verified

  • apps/docs bun run build from the worktree: 52 pages (was 50 on main: +/product/in-numbers/ +this handover’s session page), exit 0; /product/in-numbers/ in dist/, nav entry rendered; in-page canonical links rewritten (/architecture/generative/, /architecture/physical/, /roadmap/). External-request grep over dist/: no external <script>/<link>; the new page’s external URLs are pricing-source anchors only.
  • Repo root from the worktree: bun test 154 pass, 0 fail (no flake occurrence this run); bunx varlock run -- bun run build exit 0.

Hot files

None — no new dependencies; bun.lock, root package.json untouched.

Open items / next sessions

  • The page is a dated snapshot: prices change, and the C11/total rows grow once this session ends — refresh via the appendix method (foreman review is expected to re-run the tally on a sample transcript anyway).
  • C12 (docs/scope-split) plans an NF-15 → in-numbers pointer; the canonical-links rule for in-numbers.md is already in place for it.
  • If a future session adds users/telemetry, scenario (b)’s assumptions (N=10 msgs/day etc.) should be replaced with measured rates.
  • After merge: foreman dispatches deploy-static-sites.