TL;DR

# TL;DR

1. **Two variable lines move the margin: inference and research.** Infra (~$2–6/mo), Telegram delivery (**free**),
   embeddings (~cents), images (~$0.04/img) all round to noise. The lines that matter are **LLM tokens** (heavy
   creative) and the **research-loop portfolio** (Firecrawl + interpret), which on a living-domain seat (~$2.86/mo)
   is the single largest variable cost — and it **compounds** with every topic, site, and cadence the user adds (§3b).
2. **Two findings from the code that change the picture vs. the manifesto:**
   - **The performance tier IS routed now.** Under the shipped `balanced-smart` profile,
     `generate` → **performance** (Gemini 3.6 Flash, **$7.50/1M out**) — the one category that leaves the
     balanced tier. *(This corrects an earlier note here claiming `PERFORMANCE_MODEL` was assigned to no
     role; that was true of the old role-constant routing, not of `modelFor(category)` under profiles.)*
     Arc generation is the only call that reaches it, and it is **one batched call per merge**, so the
     premium tier lands on ~15% of a domain's weekly spend rather than dominating it (§3d).
   - **Output is what costs.** Output bills **5–8× input** ($2.50 vs $0.30 on balanced; $7.50 vs $1.50 on
     performance). Reading 465k tokens of articles a week costs ~$0.14; generating 27k tokens of scenes
     costs ~$0.08. The summarisation hierarchy keeps the read side cheap — the spend is on the write side.
3. **Almost every single action costs pennies.** A scene's prose is ~$0.007 COGS; a full 4-scene arc
   ~$0.005–0.03; a research tick ~$0.05 (crawl-dominated). The expensive sleepers are **TTS narration** (~$0.10/scene)
   and **Firecrawl at scale**.
4. **The model (chosen):** a **flat $20/seat that includes a generous inference allowance (~15M tokens), with
   usage-based billing on top.** The buyer sees one number; ~80% never touch the meter. Heavy users pay for what
   they pull and become *more* profitable (heavy-creative seat: **50% margin billed vs 19% if we ate it**),
   instead of forcing a throttle or a price hike on everyone. Self-host (BYO key) skips the meter. The metering
   plumbing **already exists** (`calculateApiCost`, the `GasMeter` pill) — what's left is an allowance accumulator
   + Stripe usage billing.

> **Reframing — 2026-09-20.** The Program crawl model remains valid for Directors who ground worlds in
> live evidence. The modal cost for simulation-shaped users is now generation plus Meridians World Search (MWS):
> situation intake, world construction, forked arms, readings, and debrief. The crawl model is not
> discarded; it is no longer the primary cost unit for the participant, reader, learner, or author.

---

Raw Markdown source