AI Generation Economics¶
Living document. Council-approved v2 figures (2026-07-31). Sourced figures come from the council-verified rate table in ../technical/ai-generation-strategy.md; planning figures are labelled assumptions. Plain mirror: ../plain/ai-generation-strategy.md.
This doc covers the money side of AI generation: unit economics, per-user tiers, cost at scale, top-up pack margins (WING-300), and the open decisions that move the numbers.
Unit economics (council-verified, post-promo rates)¶
| Unit | Marginal cost |
|---|---|
| Study song (Lyria bed + TTS overlay) | ~$0.13–0.17 |
| Study Cast episode (10 min, two-host) | ~$0.25–0.45 |
| Narration | $0.90–1.80 per audio-hour |
| Image (Nano Banana 2, interactive) | $0.067 ($0.034 batch, queued only) |
| Project chat turn (50k ctx, realistic caching) | ~$0.02–0.03 |
| Resume/cover-letter/doc artifact | $0.03–0.08 (Gemini Flash) / $0.10–0.25 (Sonnet) |
| Mock interview | ~$1.50–2.50 per conversation-hour |
| Grounded research turn | ~$0.01–0.035 (SKU disputed — verify) |
| Diagrams / ambient beds | $0 marginal (procedural / CC0; egress ~$0.12/GB is infra) |
Per-user tiers (generation only, per month)¶
| Tier | Cost | Notes |
|---|---|---|
| Light | ≈ $2.75 | occasional use across lanes |
| Moderate | ≈ $14.25 | the design center; doubled vs draft v1 once resume, interview, and research lanes were actually modeled — audio was never the expensive part |
| Heavy, uncapped | ≈ $57.70 | what an unlimited power user would cost |
| Heavy, capped (WING-299) | ≈ $33.10 | caps on songs, episodes, narration, chat turns, interviews |
Add ~$2–5/user for sticky IP infrastructure. Chat + mock interviews + PageAgent dominate heavy spend, so caps must cover them — not just the audio lanes.
Cost at scale¶
Mix assumption: 50/40/10 light/moderate/heavy-capped ⇒ blended generation ≈ $10.45; ~$13.95/user all-in.
| Users | Variable | Fixed infra | Total/mo | Per-user |
|---|---|---|---|---|
| 100 | ~$1.4k | ~$0.7k | ~$2.1k | $21 |
| 500 | ~$7.0k | ~$1.1k | ~$8.1k | $16 |
| 1,000 | ~$14.0k | ~$1.5k | ~$15.5k | $15.50 |
| 5,000 | ~$70k | ~$4.5k | ~$74.5k | $14.90 |
| 10,000 | ~$140k | ~$8k | ~$148k | $14.80 |
Sensitivity: switching cloud text from Claude to Gemini-primary cuts blended generation ~25–30% ⇒ 10k users ≈ $110–120k/mo. Held in reserve as a cost lever (decision log, 2026-07-31).
Top-up packs (WING-300) and margins¶
When a user hits a lane cap, they buy a pack for that lane rather than a bigger plan. Pack pricing targets ~50% margin over marginal cost — e.g. a song pack priced at ~2× the ~$0.15/song cost. Packs turn the heaviest users from a margin threat into revenue: the heavy user becomes a contributor — the ~$58 uncapped cost profile becomes ~$33 of capped cost plus pack purchases at margin, instead of a silent loss on a flat subscription. Meters are always visible; failed generations never consume quota or packs.
Promo-cliff risk¶
All math above uses post-promo Claude Sonnet 5 rates. The current promo ($2/$10 per M in/out) ends Aug 31, 2026, moving to $3/$15 — a 50% jump on the text lanes. Budgets built on promo rates would have looked ~33% better than reality; do not quote promo-rate economics anywhere.
Open decisions (each moves the numbers)¶
- Lyria indemnity verification — songs lane is blocked until Lyria is confirmed on Google's indemnified-services list; blocks the $0.15/song lane entirely.
- Suno replacement — Suno BYOK is dead (no official API). Desktop music BYOK candidates: ElevenLabs Music / Stability Audio, or drop desktop music.
- Trademark filing for the character cast (AI images likely uncopyrightable; trademark is the ownership mechanism).
- COPPA/minors policy — study features attract under-13s; age gating + data handling need a policy pass before study lanes go wide.