Atlas
kb://research/embed-tokenomics-mistral-stack-2026-06-01stable2026-06-01

OpenAlice Embed — Tokenomics recompute (full-Mistral stack) + EU/DE go-to-market

researchembedtokenomicspricinggo-to-market

OpenAlice Embed — Tokenomics recompute (full-Mistral stack) + EU/DE go-to-market

2026-06-01 · Supersedes the streaming-era cost assumptions in unit-economics-2026-05-25.md for the Embed product (that doc mixed in server-GPU kiosk rendering — Embed renders the avatar browser-side, zero server GPU). Mistral prices verified May 2026.

0. The two facts that reshape Embed economics

  1. Full-Mistral stack = EU-sovereign by default. Mistral API prices (per 1M tok, in/out): Ministral 8B $0.10/$0.10 · Small 3.1 $0.20/$0.60 · Medium 3 $0.40/$2 · Large 3 $2/$6. Batch −50%. Prompt-caching the system+brand+RAG prefix cuts input further.
  2. Voxtral (Mistral's open-weight TTS+STT, GDPR-by-design, ~10× cheaper than ElevenLabs) → the whole voice path is now Mistral/EU and cheap. Drop ElevenLabs from the default stack (keep as optional "premium voice", disclosed). Use hosted Voxtral API → no GPU on our side (Hetzner has none); self-host only if scale ever justifies renting GPU.
  3. Avatar = browser-side VRM/WebGL → €0 server render cost. No GPU line item for Embed.

1. True cost per unit (full-Mistral, hosted APIs)

UnitModel~in / outCOGS
Text convo, light (1–3 turns)Small 3.18k / 0.6k~$0.002
Text convo, typical (RAG, ~6 turns)Small/Medium blend30k / 2k~$0.01
Text convo, complexLarge 330k / 2k~$0.07
Voice add-on per voice convoVoxtral STT+TTS~1.5 min in / ~2.4k chars out~$0.06–0.10
RAG ingestion (one-time/customer)mistral-embed ($0.10/M)~5k docs~$0.50 once

Headline: text ≈ 1¢/convo. Voice ≈ 6–10¢/convo (the ONLY real cost driver). Avatar free. No GPU. → Pricing rule: be generous on text, meter voice (included minutes + overage). That holds 70–90% margin at any usage mix.

2. Tier margins on the new stack (realistic usage, voice metered)

Tier (current)PriceTypical COGS/moMarginNote
Starter€249/mo (10k convo cap)$20–60 (text)76–92%cap/meter voice; trim text cap to ~5k to protect the tail
Pro€999/mo$80–18082–92%voice overage metered; text basically free
Enterprise~€25–30k/yr (€2.5k/mo)$400–1,000/mo70–85%even heavy voice stays healthy
Pro Lifetime€9,999 once (first 30)—see ⚠️fair-use it: text-unlimited + metered voice. Unlimited voice on a pay-once plan erodes margin forever.

The €25k / lifetime offers are very profitable because the underlying cost is genuinely cents. We can present them as generous honestly. Sell voice minutes as the metered premium (sell ~€0.10–0.15/voice-min vs ~$0.06–0.10 cost).

3. NEW high-value use case: internal employee knowledge-assistant

RAG over internal docs + tools (HR/policy/IT/ops/onboarding Q&A). Per seat: ~100 queries/mo × ~$0.006 = <$1/seat/mo COGS → price €19–29/seat (min 10 seats) = 90%+ margin. Why this is the best Germany-first wedge:

  • EU-sovereignty is a *forced* buying criterion for internal company data — German firms legally can't freely pipe internal/HR/legal docs to US OpenAI. Full-Mistral = the answer almost nobody else credibly offers.
  • ROI is concrete + measurable (employee time saved). Lower brand-risk than public-facing. Text-first → tiny COGS.
  • German Mittelstand is doc-heavy, compliance-minded, privacy-conscious → ideal fit.

4. Use-case sell-map (digital now; physical later)

Ranked by ease × margin × sovereignty-wedge, EU/DE-first:

  1. Internal knowledge-assistant (above) — best margin + strongest wedge + measurable ROI. Lead here in DE.
  2. Public brand-rep widget (current flagship) — conversion ROI, flashier; higher brand-risk + voice cost. Best for innovation-forward SMB/DTC.
  3. Support-deflection (SaaS/e-com, text-first) — easy ROI math (tickets deflected).
  4. Docs/helpdesk assistant for software cos (RAG over docs).
  5. Professional services (law/tax/consulting) — doc-heavy + strong sovereignty need.
  6. Education / training — avatar tutor, onboarding.
  7. Real estate / automotive / dealerships — lead-qualification avatar (Mittelstand sectors).
Defer: physical kiosks / TV-tablet embodiment — org-side hardware capex, longer cycle. Revisit once digital revenue is flowing. Stay digital first (matches NAO's read).

5. Who's most ready (early adopters), DE/EU-first

  • German digital agencies → channel/white-label reseller. One agency = many end-clients. Fastest leverage for a solo founder.
  • DACH SaaS / tech scale-ups (Series A–B, CX-conscious, AI-curious, budget).
  • DTC / e-commerce brands (conversion-hungry, try novel UX).
  • Mittelstand with doc-heavy internal ops (HR/compliance/IT) → the internal-assistant play (slower, high-margin, sovereignty-forced).
  • AI-curious DACH founders / indie SaaS → cheap design-partner pilots (proof, not revenue).

6. Where to find them (Germany/EU, concrete)

  • LinkedIn (DACH B2B lives here) — Heads of CX/Digital/HR + founders. Send a personalized Loom of *their* site/use-case with the avatar already on it. German language.
  • OMR (Hamburg) — biggest DACH digital-marketing community/festival (agencies + brands). Bits & Pretzels (Munich), local AI/startup meetups, Web Montag.
  • Agency networks: German web/Shopify/TYPO3 agencies (TYPO3 huge in DACH) → white-label pitch.
  • Communities: IndieHackers DACH, t3n/OMR audience, German AI/SaaS Slacks/Discords, Verbände (industry associations) for vertical entry.
  • Build-in-public (planned Weekly Build Log) in German → inbound.
  • Founder-led, local, German-speaking, EU-sovereign is a real trust edge vs US tools. Coffee demos in person.

7. Fast-monetization plan (30/60/90)

  • Days 0–30: Pick the lead wedge per prospect (internal-assistant for sovereignty-forced firms; widget for agencies/DTC). Stand up a Founding Design-Partner program: 5–10 EU/DE pilots, white-glove setup + steep founder discount (€0 for 1–2 anchor logos) in exchange for testimonial + case study + feedback. 90-day "we make it work" guarantee (already in sales scripts). 5–10 personalized DMs/day.
  • Days 30–60: First 3 pilots → case studies (real numbers, replace the placeholder Ns). Pitch 5–10 German agencies a white-label/reseller deal (multiplier). Turn proof into paid mid-market deals.
  • Days 60–90: Two live motions — self-serve SMB (Mollie, €99–499/mo) + high-touch founding/enterprise (€9,999 lifetime / €25–30k enterprise annual, done-for-you). Feed CAC/conversion data back into pricing A/B.

8. Action items

  • [ ] Update pricing-page.md + landing: default voice = Voxtral (EU), drop ElevenLabs from the sovereign default (optional premium only).
  • [ ] Add per-voice-minute metering + included-minutes to every tier (voice = the cost lever).
  • [ ] Fair-use cap the Lifetime tier (text-unlimited, voice metered).
  • [ ] Add an Internal Knowledge-Assistant SKU (per-seat, €19–29, min 10 seats) — new landing section / pitch.
  • [ ] Validate per-convo token assumptions against 3 real pilot transcripts; enable Mistral prompt caching on the system+brand+RAG prefix.
  • [ ] Tycho: confirm hosted Voxtral path (no GPU our side) in the voice stack.