Implicit Prefix Cache TESTED ? Automatic, free. Gemini bills cached prefix tokens at 10% of the uncached rate (90% discount, verified April 2026 on ai.google.dev/gemini-api/docs/pricing). Validated in logs: `cachedContentTokenCount` fires on text chat turn 2+ when the RAG prefix ≥4K tok. Cache savings scale with the actual post-digest/vertex prefix share — so when digest is on, the absolute cache benefit shrinks (prefix is smaller) but percentage discount is the same. Applies to text chat + voice input. Only kicks in turn 2 onwards; turn 1 is full price.

Explicit Cache (text) PARKED · TO TEST ? Parked — code and admin panel kept in-repo but not deployed. Why: on Flash-Lite text the saving is ~$0.03/user/mo, but Gemini charges $1/M tok/hour storage while a cache is live. A 22K-tok RAG parked for 1h/show costs ~$0.022/hr — on 100 hot shows that's ~$1,600/mo of storage. Break-even needs hundreds of text turns per show per hour, unrealistic on a voice-first B2C mix. Revisit if text volume per user rises or if we move to Flash / Pro text models.

Voice Digest (voice input) TESTED ? Pre-condensed RAG used only for voice. Admin generates a digest from the full show RAG; voice session swaps the heavy episode+general+source blocks for the digest. Measured savings depend on source size and strictness: 22K RAG → ~11K normal (50%), ~7K strict (70%), ~5K ruthless (77%). Small RAGs (<6K) barely compress — digest floor is ~2–4K.

Cheap model for text chat TESTED ? ON = force Flash-Lite for text chat (overrides the Text Model dropdown). OFF = use whatever the dropdown is set to. Flash-Lite is ~2× cheaper on input and ~2× on output vs Flash 2.5, and ~16× cheaper on input vs Pro — turning this OFF with the dropdown on Flash 2.5 (default) adds roughly 30–50% to total text spend once cache discounts are blended in. Validated in production: consumer text chat ships with Flash-Lite by default.

VertexDB/RAG retrieval ESTIMATE · PENDING ? Not built yet — purely modeled. Retrieve only top-k relevant chunks from a vector DB per turn. Flat ~60% cut on the RAG share when used alone (top-k retrieval typically compresses a 20K RAG down to 3-5K) (not user-count scaled — that amortization belongs to infra cost, not per-query tokens). VOICE + DIGEST COMBO: digest provides the compressed ambient baseline (persona, world, season arcs); Vertex adds ~8% of RAG share back as targeted episode/scene retrieval (3–5 chunks × ~300 tok) — more informed voice at a small cost bump vs digest alone, still far below full RAG. Capped so the combo never exceeds full-RAG cost. On text chat (digest doesn't apply), Vertex cuts the RAG share of input directly.