Competitor Dossier — Avatar / Digital-Human Platforms

Why this exists. OpenAlice's #1 differentiator is an embodied 3D VRM avatar (three.js / @pixiv/three-vrm, voice-synced, clothing/logo-customizable) running ON a full conversational platform — analytics, HITL inbox, A/B, proactive triggers, lead capture, knowledge base — and shippable EU-resident / self-hosted. This dossier maps the avatar/digital-human market so we never re-research it. The recurring finding: almost every competitor is an avatar *engine* or *API*, not a *platform* — they render a talking face and tell you to "build the rest yourself via webhooks." That missing platform layer is the core of our moat. Read the closing section first if you only have two minutes.

Scope note / honesty: All pricing and product claims verified against official pages (June 2026) where possible; vendor-marketing numbers (engagement %, "first/only", big-logo lists) are flagged. Valuations/ARR from secondary sources are flagged. Where a source conflicts, the more recent / official figure is used and the conflict noted. Full source URLs at the end of each section.

The one distinction that organizes this whole market:

• Generation (async) — you give a script/photo/audio, you get a finished video file back. Not interactive. (Synthesia, Hedra core, HeyGen video product.)
• Real-time conversational (interactive) — a live avatar you talk to, sub-second turn-taking, WebRTC. (Tavus CVI, HeyGen LiveAvatar, D-ID V4 Agents, Simli, Anam, Hedra Live Avatars, UneeQ, Soul Machines.)
• Avatar engine/API vs. conversational platform — orthogonal to the above. Almost everyone in the real-time camp is an engine/API (render the face, BYO LLM/TTS/logic/ops). The platform layer (analytics, HITL inbox, A/B, triggers, lead capture, managed KB/RAG, embed widget with dashboard) is where the gap is — and where OpenAlice lives.

1. Tavus — Phoenix-4 / CVI (the technical SOTA in real-time)

Positioning & scale. "The human computing company" — developer-first API/infrastructure for real-time photorealistic "AI humans" via its Conversational Video Interface (CVI). Founded 2020, Houston TX (CEO Hassaan Raza, YC alum). NOT a packaged GTM application — it sells models + thin agent primitives. ~$64M total raised over 4 rounds (latest Series B, Nov 2025); investors Y Combinator, Sequoia, CRV. ~86 employees (Tracxn Apr 2026; some reviews say ~40 — small either way). Valuation/ARR not disclosed (unusual opacity for a Series B). Customers: Delphi, Final Round AI (1.2M practice minutes), ACTO, CareFlick, iAsk; big-logo lists (Meta/Salesforce/ByteDance) are un-case-studied → low confidence.

Pricing (verified tavus.io/pricing):

→ Real unit economic: ~$0.32–0.37 per conversational minute. Production entry is effectively the $397/mo Growth tier.

Avatar tech — the Phoenix lineage. Phoenix is the rendering model. P1 = NeRF 3D; P2 = switched to 3D Gaussian Splatting, broke the real-time barrier (enabled CVI); P3 = full-face generation; Phoenix-4 (Feb 2026) = current SOTA, hybrid Gaussian–diffusion: streaming audio → diffusion head generates motion coefficients → 3D-Gaussian-splat renderer. Output is 2D-photoreal via a 3D-splat representation (not mesh, not pure 2D video-gen). Fully generated real-time, not looped footage — 40 fps @ 1080p. Explicit emotion control across 10+ states, micro-expressions, active-listening backchannels. Custom replica from ~2 min consent video (or a single image, per marketing).

Perception (Raven) + turn-taking (Sparrow) — DETAIL (this is Tavus's real moat). CVI is a 6-layer pipeline: Raven → Sparrow → STT → LLM → TTS → Phoenix.

• Raven (Raven-0 → Raven-1) — "the first AI perception model that sees and reasons like a human." Continuous visual understanding: reads user expressions (confusion, smiles), gaze direction, objects, background, ambient environment, on-screen content in real time and adjusts responses dynamically — a "perception-to-expression loop" feeding Phoenix-4's emotion control. Caveat: Raven-1 trained primarily on English; emotion-perception drops to ~68.75% out-of-domain, plus EU-AI-Act risk on workplace emotion detection. So "emotional intelligence" is real but not bulletproof.
• Sparrow (Sparrow-0 → Sparrow-1) — transformer turn-taking that predicts conversational floor ownership at the frame level instead of fixed silence thresholds. On each ~100ms pause it evaluates utterance content (breathing vs. true turn-end) → confidence score → respond or wait. Multilingual via semantic context. Claims ~610ms response, "~2× faster" than fixed-threshold, interruption-aware. (+50% engagement / +80% retention are vendor marketing.)

Latency. Headline sub-600ms end-to-end (best case ~500–600ms) via stream-first WebRTC; one independent review corroborated sub-500ms. Credible, "best-case" caveat applies. This is the avatar-quality / latency bar to beat in real-time conversation.

Voice. BYO + bundled — TTS via Cartesia (default), ElevenLabs, or Azure; bring any OpenAI-compatible LLM. No proprietary-TTS lock-in.

Platform layer — THE GAP.

• Included (API-configured primitives): Knowledge Base / RAG (upload PDF/CSV/PPTX/URLs, ~30ms retrieval), cross-session Memory, Guardrails & Objectives, Function Calling, a raw "conversation data layer" (transcripts, emotion timelines, perception events, sentiment).
• MISSING (build-it-yourself via API/webhooks): ❌ HITL inbox / live takeover (escalation is just a function-call trigger, no human console) · ❌ A/B testing · ❌ proactive/outbound triggers (CVI is inbound, session-initiated) · ❌ lead-capture forms · ❌ a real analytics/BI dashboard product (only raw data via API) · ❌ no-code/visual flow builder · ❌ no official MCP server.

Customization. Custom replicas (~2 min video) + 100+ stock; custom persona/guardrails; backgrounds via green-screen mode (client-side DIY composite, not a built-in picker). No clothing/outfit swap found. 100% white-label on Enterprise.

Channels. Web embed (@tavus/cvi-ui React + iframe); WebRTC video call (Daily default, or LiveKit). ❌ No native phone, ❌ no packaged kiosk.

Multilingual. ~42 (docs) to "50+" (marketing) languages.

Security/compliance. SOC 2 Type 2, HIPAA, GDPR confirmed (Trust Center). EU data residency: unverified (ask sales). Consent: requires a consent video to train a replica; detailed anti-deepfake mechanics not publicly documented → unverified. EU-AI-Act emotion-detection risk flagged.

Self-host. ❌ No self-host / on-prem / VPC. Managed cloud API only — they "handle the real-time infrastructure." A genuine Tavus weakness for private/EU buyers.

API/SDK. REST (OpenAPI), TS/JS/Python; React @tavus/cvi-ui; Daily / LiveKit transports; first-class Pipecat (currently the only video provider for Pipecat); webhooks. No MCP server.

Verdict. Tavus is the technical SOTA of real-time conversational avatars (Phoenix-4 + Raven + Sparrow + ~600ms) — and the clearest proof of the market pattern: it is infrastructure, you bring the application. No HITL, no A/B, no triggers, no lead forms, no dashboard product, no self-host. This is exactly the layer OpenAlice supplies.

Sources: tavus.io · tavus.io/pricing · tavus.io/cvi · docs.tavus.io/.../overview-cvi · tavus.io/blog/phoenix-4-real-time-human-rendering-with-emotional-intelligence · tavus.io/blog/sparrow-0-... · businesswire 20260218278213 (Phoenix-4) / 20250306296766 (Raven-0/Sparrow-0) / 20260120774715 (Sparrow-1) · marktechpost.com/2026/02/18/... · chatforest.com/reviews/tavus-conversational-video-ai-api · tracxn.com/d/companies/tavus · crunchbase.com/organization/tavus

2. HeyGen — Interactive Avatar / LiveAvatar + video-gen leader

Positioning & scale. Two pillars: (1) async AI talking-head video generation (photoreal "digital twins" from a photo, 175+-language translation, voice cloning, Video Agent) — the mature revenue product; (2) real-time avatars = LiveAvatar (rebrand of beta "Interactive Avatar"), a WebRTC streaming-avatar API now on a separate domain (liveavatar.com) — avatars not portable between the two products. Series A: $60M, Jun 2024, led by Benchmark, at $500M valuation (confirmed, Bloomberg). Series B reported ~$150M @ ~$2.0B (mid-2025) — secondary only, unconfirmed. ~$95–100M ARR by late 2025 (most consistent figure); 1M→$100M ARR in ~30 months with <200 staff. Employees ~200–360 (post-Genova-Labs acquisition, Sept 2025). Customers: Workday, Würth; "70% of Fortune 500 L&D" is unverified marketing.

Pricing.

• Subscription (video): Free $0 · Creator $29/mo (600 credits, 175+ langs) · Pro $49 · Business $149/mo + $20/seat (SSO, SCORM) · Enterprise custom. Note: Avatar IV ≈ 20 credits/min, so Creator buys only ~30 min premium video/mo.
• API (PAYG, no tiers): $1 = 1 min standard avatar video; Avatar IV $4/min; translation/Video Agent $2/min; 10 concurrent; no free API credits (Feb 2026).
• Real-time (LiveAvatar): credit = $0.10. Full mode ≈ $0.20/min (HeyGen runs whole ASR+LLM+TTS), Lite mode ≈ $0.10/min (BYO stack). Plans: Starter $19/mo (5 concurrent, 720p) · Essential $99 (20 concurrent) · Business $475 (40 concurrent, 1080p) · Enterprise up to 100 concurrent.

Avatar tech. 2D photoreal talking heads (not 3D). Flagship async model Avatar IV (May 2025): single still photo → talking avatar, "diffusion-inspired audio-to-expression engine," handles off-axis poses, Fast/Quality modes. Real-time LiveAvatar needs a 2-min consent capture (listen 15s / talk 90s / listen 15s) — no instant-from-photo real-time avatar.

Lip-sync. Reviewers rate HeyGen above Synthesia on micro-expressions, gestures, eye-line, "production-grade lip-sync." No published benchmark — reviewer opinion + marketing.

Latency. No official ms figure. Marketing says "low latency / fastest on market." A secondary blog cites a sub-200ms goal for late 2026 (i.e., not there yet, aspirational). Community reports significant lag when a custom KB is wired in. Full mode adds hops (their ASR + GPT-4o-mini + ElevenLabs + render).

Voice. Not fully in-house for real-time — LiveAvatar uses ElevenLabs Flash v2.5 (TTS), Deepgram/AssemblyAI (ASR), GPT-4o-mini (Full-mode LLM). Voice cloning in 175+ langs (consent + verification required). "The magic is partly rented."

Platform layer — GAP. LiveAvatar is fundamentally API + Web SDK, "build the rest yourself." KB/RAG = thin / bring-your-own (and causes lag); connect-your-own-LLM ✅ (Lite). ❌ no first-party analytics dashboard product · ❌ no HITL inbox / takeover (absent) · ❌ no A/B · ❌ no proactive triggers · lead capture = a use-case pattern you implement, not a module. An engine, not an operations platform.

Customization. Custom digital twins (1–5 by plan), unlimited photo avatars, 500+ stock; no in-app outfit/clothing swapper confirmed (look = source capture). Brand voice via cloning.

Channels. Web embed via LiveAvatar Web SDK (you build the UI); async export 720p/1080p/4K, SCORM/LMS (Business). REST API + Streaming SDK (@heygen/liveavatar-web-sdk), WebRTC.

Multilingual. 175+ languages (translation/voice clone headline); real-time langs constrained by TTS/ASR providers (else Lite + your own stack).

Security. SOC 2 Type II, GDPR, CCPA, EU-US DPF, EU-AI-Act; TLS 1.2+, AES-256, no training on customer data. Hosted AWS-US; data stored in US; no advertised EU residency.

Self-host. ❌ None. SaaS-only on AWS-US. Hard blocker for private/EU-resident buyers.

Verdict. Best-in-class photoreal rendering + strongest brand/ARR/funding; real WebRTC streaming with a clean SDK. But no conversational-ops layer, no published latency, no self-host, US-only residency, realtime depends on third parties, fragmented LiveAvatar split. Engine, not platform.

Sources: heygen.com/pricing · heygen.com/api-pricing · help.heygen.com/.../heygen-api-pricing-explained · help.heygen.com/.../introducing-liveavatar · /liveavatar-faq · liveavatar.com · docs.liveavatar.com · heygen.com/avatars/avatar-iv · heygen.com/blog/announcing-our-series-a · bloomberg.com/.../heygen-valued-at-500-million · sacra.com/c/heygen · heygen.com/security · github.com/heygen-com/liveavatar-web-sdk · anam.ai/blog/heygen-api-...

3. D-ID — Agents / Clips (the one with a real platform layer)

Positioning & scale. Tel-Aviv (founded 2017), pivoted from photo de-identification to "Creative Reality." Now leans hard into real-time conversational visual agents for support/sales/training/kiosks. ~$48M total raised (7 rounds; last major = $25M Series B led by Macquarie Capital) — ~1/10th of Synthesia. Valuation not disclosed. Acquired simpleshow (explainer video, Dec 2025). Cites 1,500+ enterprise customers, millions of subscribers. No ARR.

Pricing. Credit model (1 credit ≈ 15s video; streaming credits halved). Studio: Lite $5.90/mo (10 min, watermark) · Plus ~$16 · Pro ~$48 (API) · Advanced ~$80–299 (tier names/prices inconsistent across reviews) · Enterprise custom. API headline ~$5.90/min (streaming ≈ half). Agents billed on speaking time: ≤15s message = 0.5 credits.

Avatar tech. Both pre-rendered (Talks/Clips) and real-time. Generations: V2 (instant photo→talking head, "Live Portrait"), V3 Instant (<10 min input), V3 Pro (3–5 min capture, 24h build), V4 Expressive Avatars (multi-emotion recordings). Dual generation+real-time is D-ID's key differentiator vs Synthesia.

D-ID Agents — a genuine platform layer (rare in this market):

• ✅ Knowledge base / RAG — upload PDF/TXT/PPTX + website URLs (max 5 docs/agent, 500K chars each), RAG retrieval.
• ✅ Embed widget — drop onto site/app/LMS/portal/kiosk "in seconds."
• ✅ Live voice / video-call interaction (type or mic).
• ✅ Analytics — built-in engagement tracking.
• ✅ Triggers/actions — workflows, fetch data, book meetings, V4 MCP Apps render inline UI (charts/forms/quizzes).
• ✅ Connect any LLM (model-agnostic).
• ❌ No HITL inbox / human handoff, ❌ no lead-capture forms, ❌ no A/B (agent-config-centric, not CX-ops-centric).

Lip-sync & latency. V4 (16 Mar 2026): sub-0.5s turn latency, up to 4K, diffusion-based, sentiment-aware, optional camera layer to read user nonverbals. Lip movement within ~30ms of speech output (third-party). Earlier: ">90% answer accuracy in <2s."

Voice. Azure + Amazon Polly (long-standing standard) + ElevenLabs Pro voices + voice cloning + native-language voices.

Customization/channels/multilingual. Embed widget, API, RTMP, kiosks, LMS, mobile. 120+ languages/accents (100+ with lip-synced translation).

Security. SOC 2 + ISO 27001 / 27017 / 27018 + GDPR (SCCs); strong consent/anonymization heritage. Runs on Azure. EU residency / self-host: not affirmatively documented (inferred-absent; Enterprise may negotiate).

Verdict. D-ID is the closest direct analog to OpenAlice Embed on the live conversational-avatar + embed-widget + KB/RAG + analytics axis — and the only competitor here with a meaningful built-in platform layer. But: still 2D photoreal (not embodied 3D), and lacks the CX-ops layer (HITL inbox, lead capture, A/B) plus no documented self-host / EU residency. They are the platform-layer benchmark to out-build.

Sources: d-id.com/ai-agents · d-id.com/ai-avatars · d-id.com/news/v4-expressive-visual-agents-... · d-id.com/pricing/api · d-id.com/pricing/studio · help.d-id.com/.../how-does-agent-pricing-work · d-id.com/security-and-privacy-compliance · d-id.com/blog/d-id-achieves-soc-2 · /d-id-has-achieved-iso-certification · d-id.com/blog/d-id-closes-25-million-funding-round · computerweekly.com/.../inside-d-ids-real-time-ai-avatar-technology

4. Synthesia — enterprise video-gen unicorn (NOT real-time)

Positioning & scale. "The #1 AI Video Platform for Business" — enterprise L&D / training / internal comms (UK). $200M Series E (26 Jan 2026) at $4.0B valuation, led by GV (Google Ventures), with Kleiner Perkins, Accel, NEA, NVIDIA NVentures, PSP Growth, Hedosophia, FirstMark. Total raised ~$530M; valuation doubled $2.1B→$4.0B in a year. ARR crossed $100M (Apr 2025); ~$146M Sept 2025 (Sacra est.). ~600 employees (the "783" figure unverified). 60K+ business customers, 1M+ users, used by 90%+ of the Fortune 100. Customers: SAP, Bosch, Merck, Heineken, McDonald's.

Pricing. Free $0 (10 min/mo, watermark) · Starter $29/mo (120 min/yr) · Creator $89/mo (360 min/yr, API + interactive video) · Enterprise custom (unlimited, SSO/SAML, SCORM, 80+-lang translation, Video Agents rolling out). Minute-metered.

Avatar tech & the real-time question (CRITICAL). 240+ stock avatars; personal/custom from a photo or short video. Core product is PRE-RENDERED: text→video, output is a downloadable/shareable file. Latest model = Express-2 (Synthesia 3.0, Oct 2025): diffusion-transformer, full-body avatars with hand/body gestures, micro-expressions, "perfect lip sync," 1080p/30fps — still pre-rendered. (No "Avatar 4/5" product line — that naming is wrong; the line is Expressive Avatars → Express-1 → Express-2.) The "Video Agents" feature (markets "talk/listen/act in real time") is waitlisted, Enterprise-only, scoped to internal training role-play (practice + scorecard) — NOT customer-facing live chat. Synthesia is unambiguously an async video generator, not a real-time conversational platform.

Voice. Proprietary Express-Voice (in-context cloning "in seconds") + 1,000+ AI voices; 140+ languages (some pages say 160+).

Platform layer — content/LMS, NOT conversational. Deep content platform: workspaces, roles, real-time co-editing, guest seats; video analytics (views/drop-offs/completion, incl. SCORM/embed); brand kits, branded pages; SCORM 1.2 & 2004 export (auto-propagating updates); AI dubbing + 1-click translation (140+ langs, lip-synced); Interactivity 2.0 (quizzes/polls/hotspots/branching — interactive pre-rendered video, not live conversation). Conversational layer CONFIRMED ABSENT: no HITL inbox, no human handoff, no lead capture, no message/keyword triggers, no live-chat routing, no CRM conversation history. "Engagement" = "did they watch the video," not a conversation.

Security. SOC 2 Type II, ISO 27001, ISO/IEC 42001 (claims first gen-AI co. certified to the AI-management standard), GDPR. EU data residency (Ireland DCs; AWS Frankfurt backup). AES-256/TLS 1.2+, SSO/SAML, C2PA / Content Authenticity Initiative member, consent required before cloning. No on-prem (cloud SaaS, EU-hosted).

Verdict. A $4B content/L&D powerhouse with a deep video/LMS platform — but a different category: zero real-time conversational layer. Not a head-to-head competitor for live embedded avatars; relevant as the enterprise-trust + EU-residency + compliance bar (SOC2/ISO27001/ISO42001/C2PA/EU residency) we should match for enterprise sales.

Sources: synthesia.io/post/series-e-200-million-4-billion-valuation · techcrunch.com/2026/01/26/synthesia-hits-4b-valuation · sacra.com/c/synthesia · synthesia.io/pricing · synthesia.io/features/avatars · synthesia.io/features/video-agents · synthesia.io/post/express-2-... · synthesia.io/legal/security-practices · synthesia.io/post/synthesia-is-iso-27001-certified

5. Hedra — Character-3 / Omnia (generation studio, not a platform)

Positioning & scale. A generative-media studio built on in-house omnimodal foundation models (image+text+audio→video). $32M Series A led by a16z (May 2025) at ~$200M valuation; ~$43–44M total. Investors: a16z (Infrastructure + Speedrun), Abstract, Index. CEO Michael Lingelbach. Self-reports 3M+ users / 10M+ videos (marketing, unverified). Broadening into a multi-model studio (~28 models: Kling, Veo, Sora, Flux).

Pricing. Credit subs: Free $0 (100 cr) · Basic $15 · Creator $30 · Pro $75 · Teams $75 · Enterprise custom. Credit burn varies by model. Platform API (Feb 2026) draws from the same balance. Live Avatars (real-time) billed separately at $0.05/min (~$3/hr).

Avatar tech — the key distinction. Two product lines: Character-3 / Omnia = GENERATION-ONLY (async: supply image+audio/prompt → get a finished video; Omnia, Feb 2026 is the flagship — full-body + face from one image; Character-4 is NOT shipped, "still cooking"). Live Avatars = real-time but a separate, bolted-on feature (Jul 2025) built on LiveKit, sub-100ms video response, BYO LLM (OpenAI/Gemini/Claude) + TTS (ElevenLabs/Cartesia). Reputation: best-in-class image-to-talking-video fidelity, but real-time latency rated "Slow" vs Simli/Anam in third-party tests.

Platform layer — ABSENT. No analytics, no HITL, no triggers, no KB, no conversation orchestration. Live Avatars = a face layer; you bring the whole agent stack. No self-host, no EU residency, no published compliance posture.

Verdict. Competes on generation fidelity + a16z credibility, NOT as a conversational platform. Their core is generation; real-time is a thin LiveKit face layer.

Sources: a16z.com/announcement/investing-in-hedra · globenewswire.com/.../hedra-raises-32m · siliconangle.com/2025/05/15/... · hedra.com/pricing · medium.com/@CherryZhouTech/hedra-live-avatars-... · mpost.io/hedra-launches-live-avatars-with-livekit · x.com/BrentLynch/status/1978844135978381650 (Character-4 "still cooking")

6. Shorter notes — Soul Machines, UneeQ, Yepic, Simli, Anam.ai

Soul Machines (soulmachines.com) — ⚠️ IN RECEIVERSHIP. The premium 3D "Digital People" pioneer (biologically-inspired autonomous animation, "Human OS"), real-time emotionally-expressive 3D for enterprise CX. Raised $135M+ (Horizons Ventures, Temasek, Mercedes-Benz). Collapsed: headcount 253 (Jul 2023) → ~45 → voluntary receivership 5 Feb 2026 (KPMG NZ, seeking buyer, owing $19.6M+); founders Cross/Sagar already departed; Mercedes/ANZ/Air NZ exited. Do not treat as a stable competitor — the canonical cautionary tale that high-fidelity real-time 3D ≠ a sustainable business without the platform/economics around it. (Real-time 3D: yes. Self-host/EU: unclear/moot. Platform: yes historically.)

UneeQ (uneeq.com / digitalhumans.com). Enterprise 3D digital-human conversational platform (brand ambassadors, support, kiosk+web) via its Synanim animation engine, partnered with NVIDIA ACE. Founded 2009 (Austin; NZ roots, ex-FaceMe). Lightly funded (~$10M) yet still operating, claiming record growth late 2025 (PwC, Middle East). The pragmatic survivor in 3D enterprise — and notably a genuine conversational *platform (their differentiator vs pure APIs). (Real-time 3D: yes. Self-host/EU: not documented. Platform: yes.)* Closest conceptual analog to OpenAlice's embodied-3D + platform combo — but 3D is mesh/rig (not VRM), heavyweight enterprise, no self-host/EU story, no streaming-product breadth.

Yepic AI (yepic.ai). Real-time 2D avatar API, value-positioned (from ~$20/mo); function-calling, embed, concurrent sessions, 4K downloads. Controversy: backlash over curtailing legacy "lifetime" deals (reputational liability), mixed Trustpilot. Small/scrappy, funding undisclosed. (Real-time 2D: yes. Self-host/EU: not documented. Platform: partial.)

Simli (simli.com) — YC, infrastructure-first. Low-latency 2D streaming face API — explicitly "the face layer for AI agents," not a studio. Founded 2023, YC-backed (amount unverified). Sub-300ms speech-to-video; new "Gaussian models" renderer. Cheapest in market (~$0.009–0.05/min; free tier). Pure API, BYO agent stack. (Real-time 2D: yes. Self-host/EU: no. Platform: no — face layer only.) Strength: latency + price.

Anam.ai (anam.ai) — UK/EU, developer-first. Real-time conversational 2D avatar API ("emotive personas"), London, founded 2024. SDKs (JS/Python/REST), 70+ languages, regional session deployment (soft EU-residency story, no hard guarantee). Latency ~180ms model / sub-1s median (rated "Good"). ~$11.3M raised (£1.75M pre-seed Concept Ventures + $9M seed Jul 2025, Redpoint + SV Angel; angels incl. ElevenLabs CEO). Pricing: free 30 min/mo; $0.16→$0.11/min (Starter→Pro), per-second billing. (Real-time 2D: yes. Self-host/EU: soft EU via regional deploy, no hard residency/self-host. Platform: mostly avatar-API + conversation engine, not full analytics/HITL/KB.) The most credible EU-leaning low-latency 2D-API competitor — but still API, not platform, and 2D not embodied 3D.

7. Comparison Table

8. OpenAlice's Moat — embodied avatar × full platform × EU self-host

The triple nobody else has. Map every competitor onto three axes and the gap is structural, not incremental:

1. Embodiment (3D rigged VRM, not 2D photoreal video). Everyone real-time-credible here is 2D photoreal talking-head video (Tavus's splat-based, HeyGen, D-ID, Hedra, Simli, Anam). The only 3D players are Soul Machines (in receivership) and UneeQ (mesh/rig, heavyweight enterprise). Nobody ships a real-time, voice-synced, *VRM*-based embodied avatar with clothing/logo customization on three.js/three-vrm. Our embodiment is a different medium — controllable rig, brandable wardrobe, stylized-or-real, cheap to render, ownable — not a rented photoreal face.

1. Full conversational platform (the layer 90% of them explicitly tell you to build yourself). Tavus, HeyGen, Hedra, Simli, Anam are avatar engines/APIs — "build the rest via webhooks." Their own pages confirm: no HITL inbox/takeover, no A/B, no proactive triggers, no lead-capture module, no first-party analytics dashboard product. Only D-ID (2D) and UneeQ (3D enterprise) bring a meaningful platform layer — and even D-ID lacks the CX-ops layer (HITL inbox, lead capture, A/B). OpenAlice ships the whole operating surface: analytics, HITL inbox, A/B, proactive triggers, lead capture, managed RAG/KB, embed widget with a dashboard (openalice-embed / openalice-embed-app), plus its own voice (openalice-tts/stt), realtime (openalice-rt/realtime), animation (openalice-anim), vision (openalice-vision), and RAG (openalice-rag) stack — not rented from ElevenLabs/Deepgram/GPT-4o-mini.

1. EU residency + self-host (almost nobody). Tavus, HeyGen, Hedra, Simli — no self-host, no/unverified EU residency. HeyGen is hard US-only. Synthesia has EU residency but no self-host and no real-time. Anam is a soft EU story (regional deploy, no hard guarantee). No competitor offers genuine self-host / on-prem / VPC for a real-time conversational avatar. OpenAlice has a dedicated openalice-selfhosted path and EU hosting — a category-defining wedge for regulated/government/EU-data-sovereignty buyers that the entire field structurally lacks.

The intersection = empty. No single competitor sits in (embodied 3D) ∩ (full conversational platform) ∩ (EU self-host). Tavus owns avatar tech but not the platform/self-host. D-ID owns the most platform but is 2D/cloud-only. UneeQ is 3D+platform but no self-host/EU/streaming-breadth and is enterprise-heavy. Synthesia owns enterprise trust+EU but isn't real-time. OpenAlice is the only one targeting all three at once. That triple is the moat — and it compounds with our streaming/products surface (Embed, World, Social, Persona, Live) that none of them have.

The avatar-quality bar we must push toward. The moat is structural, but the avatar/perception quality bar is set by Tavus and we should close the gap deliberately:

• Latency: target sub-600ms utterance-to-utterance (Tavus's verified bar; Simli sub-300ms and Anam ~180ms model latency show 2D can go lower — our 3D rig should aim ≤600ms to be competitive, ≤300ms to lead).
• Turn-taking: match Sparrow-class semantic, frame-level turn-taking (content-aware end-of-turn detection, interruption handling) — not fixed silence thresholds. This is a concrete, copyable target.
• Perception: Raven-class visual perception (read user expression/gaze/ambient via webcam, feed it back into avatar emotion) is the frontier feature — high-value, and Tavus's own English-only / 68.75% out-of-domain weakness is an opening to do multilingual perception better.
• Lip-sync / expressiveness: drive VRM blendshapes from audio at Phoenix-4-class fidelity (10+ emotional states, micro-expressions, active-listening backchannels, gaze/blink). Our advantage: a rig gives deterministic, controllable, brandable expression vs. a generative face — lean into that rather than chasing photoreal pixel-parity.
• Enterprise-trust bar (orthogonal but required to *sell* the moat): match SOC 2 Type II + ISO 27001 + ISO 42001 + C2PA/consent + GDPR (Synthesia/D-ID set this) so the self-host/EU advantage is credible to the regulated buyers it's aimed at.

One-line positioning. "Everyone else rents you a photoreal face and tells you to build the platform. OpenAlice gives you an embodied, brandable 3D avatar **and** the whole conversational platform — analytics, inbox, A/B, triggers, leads, knowledge — and you can run it in the EU or on your own metal. Tavus is the avatar-quality bar; nobody is the whole-stack bar but us."

Honesty ledger (what's soft)

• Valuations/ARR for Tavus, HeyGen Series B are secondary/unconfirmed; treat HeyGen ARR as "~$100M late-2025," Tavus valuation as undisclosed.
• "Engagement/retention %" claims (Tavus +50/+80), "first/only" superlatives, big-logo customer lists are vendor marketing.
• EU-residency/self-host absence for Tavus/D-ID is inferred (not affirmatively denied on their pages) — confirm with sales before quoting in a deal.
• Latency figures are best-case vendor claims except where a third-party corroborated (Tavus sub-500ms, Anam "Good").
• OpenAlice stack facts (three-vrm, openalice-embed/embed-app/selfhosted/tts/stt/rt/realtime/rag/anim/vision) verified by repo presence in org-openalicelabs; exact shipped latency/perception parity is a build target, not a measured claim.