Use of generative AI

Almost every "smart" thing ASMI does is a Gemini call under the hood, and this chapter is the honest inventory of them: what runs a model, which model, for what, and how that shows up on your credit balance. Two principles make the whole picture legible.

Everything routes through one chokepoint. Every Gemini call in the product goes through a single wrapper that records the token and image counts — never the prompt or response text, only counts and identifiers, so the usage log is GDPR-safe to keep — then prices the call and deducts the credits. There is no second path, which is also why the list of permitted models is short and deliberately curated.

Design-time is on me; runtime is on you. Generation you trigger while building and testing in the editor runs on Gemini and spends your ASMI credits. The live conversation of a deployed avatar does not — it runs on the host site's own LLM, billed to the host. ASMI is a design-time tool; it is not in your visitors' critical path. (The one wrinkle: the in-editor Simulator and Narrator, and the legacy hosted script-tag widget, run that conversation on ASMI's own servers — so testing a chat does spend credits. A coding-agent deployment does not.)

The models

All of them are Gemini, kept to a short allowlist (extending it is a security-reviewed decision). Three text models and the Nano Banana image family:

Every inference call, by purpose

Building the face (image generation)

Teaching it (text generation)

The conversation (spends credits only in design-time / testing — see above)

Gallery and safety

One non-avatar pipeline shares the same chokepoint — drafting a LinkedIn post for the Broentech blog — but that's the site's admin tooling, not ASMI.

And a note on what is not a model call, because it's easy to assume otherwise: brand-voice compilation — turning the seven sliders, the persona, and the free-form rules into a system prompt — is deterministic string assembly, no inference. The state machine's guards and transitions are pure logic too. Only the rows above hit a model.

How inference becomes credits

A credit is ASMI's unit of metered AI work. One credit corresponds to roughly $0.02 of underlying Gemini cost — the buffer that keeps the retail margin honest. The chokepoint converts every call to credits with a simple, model-aware rule:

• Text — about 1 credit per 10,000 tokens (prompt and completion combined).
• Images — a flat per-image charge that scales with the model: 2 credits for a Nano Banana frame, 4 for Nano Banana 2, 7 for Nano Banana Pro (each derived from the model's per-image price against the NB1 baseline).

Credits are deducted after the call, against measured usage, and never below zero — a cheap pre-check fails fast when you're low, and the final deduction is capped at your remaining balance.

Where credits come from, by plan:

Run low mid-month and you can top up with one-time Boost Packs (500 / 2,000 / 10,000 credits). Every plan also carries a generous hard monthly spend ceiling — a safety net against a compromised account, not a limit honest use will reach.

Two plan gates connect straight back to the model tables above:

• Model choice (Pro/Business) unlocks the auto-QA retry ladder's higher-grade rungs — Nano Banana 2 and Nano Banana Pro — for auto-healing a frame that drifts from the locked identity. Baseline frame generation runs on Nano Banana on every tier.
• Auto-consistency-QA (Pro/Business) is the consistency-QA loop in the face table: the automatic verification and premium-model auto-healing. Lower tiers still get consistency-by-construction generation, just not the automatic re-check and escalation.

And the line worth repeating, because it is the whole economic model: once an avatar is deployed, its conversations cost you nothing in ASMI credits — they run on the host site's own LLM key. Your credits are spent building and testing the avatar, not answering your visitors.