Model unit economics
Unit economics for the prototype: subscription revenue and inference spend for one primary model. Inputs are assumptions, not measurements — calibrate against real Gemini promptTokenCount /candidatesTokenCount logs as they accumulate.
Monthly P&L
Caching only applies to the system-prompt portion of input. Gemini cached rates assume a paid Vertex AI / AI Studio tier. Anthropic cached rates use the published 90%-off cache-read price; the one-time 25% cache-write surcharge isn't modeled — assume the cache is hot.
Models the Flash → Pro fallback. Cost is the escalation model's per-call rate × calls × escalation rate. If the escalation model equals the primary, this row stays at $0 — no double bill.
Every cost-bearing stage that is not the analyze call. Entered as $/clip rather than modelled from tokens because TTS bills per character and the judge is Anthropic — a token model would be fiction. All five are metered server-side and seed fromcost_by_stage. Omitting them was the largest single error this page had (#1641): on the first fully-metered prod day they summed to ~$0.030/clip against analyze's $0.044.
storage + compute fromcost_by_stage — the day's bill divided by the day's clips, the same allocation /costhatches. Not a per-clip measurement. Measured at ~$0.023/clip, against the $0.002 this page assumed before #1641. Use Reset to measured rather than calibrating by hand.
Covers what doesn't scale with calls: BigQuery billing export, scheduler triggers, dashboard reads, base infrastructure.
Actions runner minutes (Linux + macOS), Copilot seats, GitHub Packages, and shared LFS storage. Flat with respect to users — engineering overhead, not per-clip cost.
Optimal price forecast
Sweeps subscription price across the range using a constant-elasticity demand curve anchored at your current Subscription × Users:users(p) = u₀ · (p₀/p)^ε. Elasticity 1.0 means a 10% price hike costs 10% of users; 1.5 is typical for prosumer SaaS; <1 means demand is inelastic (no interior optimum — profit grows with price until the curve tops out at the upper bound). Treat the optimum as a starting hypothesis for a real pricing test, not a target.
Price history
No price history captured yet.
$/1M tokens over time from the daily pricing sync. Points before the first sync are seeded from git history and are approximate.