Tavus — The Conversational Video AI That Builds Machines That See, Hear, and Respond Like Humans in Real Time
Most AI video companies are in the content production business. They take a script, a digital avatar, and a voice model, and they produce a video file. The workflow is asynchronous: you input text, you receive video. The output is polished, usable, and fundamentally pre-recorded — even if it was generated by AI on demand.
Tavus is not in that business.
Tavus is building something they call “human computing” — the idea that software interfaces can replace the text cursor and the chatbot with a face, a voice, and a nervous system that listens, understands, and responds in real time. Not a recorded avatar playing back a script. A conversational entity that perceives tone and hesitation, models turn-taking probabilistically, generates emotional micro-expressions per frame, and speaks back in under 500 milliseconds.
The difference between async AI video and real-time conversational AI video is not a product feature distinction. It is an architectural distinction. The model stack, the latency constraints, the integration patterns, and the use cases are different in kind, not just degree.
Tavus is betting that the second category — real-time conversational video AI — is larger, harder to replicate, and more valuable over time than the first. Their proprietary research stack, their YC pedigree, and their $58M+ in funding from Sequoia and others are the evidence they are putting behind that bet.
This review covers what Tavus actually built, how it works, who it serves, what it costs, and where it is genuinely ahead of — and behind — the platforms it competes with.
We research AI tools from public sources and documentation. We do not test them hands-on.
The Founders: YC Summer 2021, a Thesis About Machines and Humans
Tavus was founded in 2021 by two people whose backgrounds converge on the same problem from different angles.
Hassaan Raza (Co-Founder & CEO) brings the commercial and product perspective. He has led Tavus through its Y Combinator batch, its Series A, and its repositioning from a personalized video tool for sales teams into a foundational infrastructure layer for real-time human-like AI.
Quinn Favret (Co-Founder & COO) provides the operational and go-to-market depth that converts research output into deployable products. Together, they have kept Tavus lean — approximately 40 employees as of 2025, a small team relative to the technical ambition.
The company is headquartered in San Francisco and emerged from the YC Summer 2021 cohort. Their GitHub organization describes their mission as “Teaching machines how to be human” — which reads like a tagline but is also a precise description of their technical roadmap. Phoenix-4, Raven-1, Sparrow-1, and Hummingbird-0 are not product names chosen for marketing resonance. They are the components of a machine perception-and-expression system: rendering, perception, timing, and lip synchronization, each built as a named research artifact with published benchmark results.
The research-first posture is characteristic of YC alumni who build infrastructure rather than applications. Tavus has published technical blog posts for each model release rather than only press announcements. Whether this reflects genuine technical depth or good developer marketing is a question that the benchmarks — and the customers — are beginning to answer.
Funding History
Tavus has raised approximately $58M across confirmed rounds:
| Round | Date | Amount | Notes |
|---|---|---|---|
| Seed | 2021–2022 | Undisclosed | Sequoia Capital confirmed as early investor |
| Series A | March 2024 | $18M | Reported by TechCrunch; platform opened to developers with face/voice cloning API |
| Series B | November 2025 | ~$40M | Reported by Axios via YC company profile; no direct press release confirmed |
Total raised: approximately $58M (USD). Valuation has not been publicly disclosed.
The Series A timing is significant. March 2024 is when Tavus moved from closed beta to open developer access with a face and voice cloning API. That was also the period when real-time AI video was transitioning from a demo category into a deployable one — latency numbers were dropping, model quality was improving, and developers were beginning to ask whether conversational video could replace text chat in specific contexts.
Sequoia’s presence as an early investor adds credibility. The firm has a track record of backing infrastructure plays early. The Series B of approximately $40M, if confirmed at its reported size, suggests the market has validated Tavus’s real-time thesis with enough capital to build a durable engineering team.
What Tavus notably lacks is a publicly disclosed revenue figure. There is no confirmed ARR, no stated user count, no published customer count. This is unusual for a company at Series B stage with $18M in their first disclosed round. The most reasonable interpretation is that Tavus has chosen to compete on technology depth rather than growth metrics — which is a viable strategy for a developer infrastructure company but creates information asymmetry for buyers and analysts.
The Technical Architecture: Four Models, One Thesis
What distinguishes Tavus most clearly from HeyGen and Synthesia is not the avatar quality or the pricing or the enterprise certifications. It is the model stack.
Tavus has built and named four distinct research systems, each solving a specific subproblem of real-time conversational video AI. This is not a vendor assembling commodity models from third-party APIs. It is a research organization building proprietary layers and publishing benchmarks for each.
Phoenix-4: Real-Time Rendering with Emotional Intelligence
Released February 2026, Phoenix-4 is Tavus’s visual rendering model. Its specifications:
- 40 frames per second at 1080p resolution
- Full-duplex operation: the model simultaneously listens and speaks, generating visual responses while audio input is still arriving
- Millisecond-level rendering latency (distinct from end-to-end conversation latency)
- 10+ controllable emotion states: happiness, sadness, anger, surprise, disgust, fear, excitement, curiosity, contentment, and additional variants
- Active listening behaviors: nodding, micro-expressions, and gaze direction generated per-frame based on conversation context — not canned animations triggered by keywords
- 3D Gaussian Splatting renderer: bypasses mesh-based constraints, generates every pixel from scratch rather than deforming a template mesh
The Gaussian Splatting approach is technically significant. Traditional avatar rendering applies geometric transformations to a 3D mesh — it is fast and controllable, but the mesh creates constraints on what expressions and angles are achievable. Gaussian Splatting is volumetric: it represents the face as a point cloud, and every frame is synthesized rather than deformed. The tradeoff is computational cost; the benefit is expressiveness.
The previous generation — Phoenix-3 — ran at 30fps at 1080p with no emotion control. Phoenix-4’s addition of 10+ emotion states represents a meaningful capability step, not just a performance improvement.
Raven-1: Multimodal Perception
Released simultaneously with Phoenix-4 in February 2026, Raven-1 is the perception layer — the component that processes what a human is saying, how they are saying it, and what their face is doing, then translates that into information downstream models can act on.
Raven-1 processes three signal streams simultaneously:
- Audio signals: tone, prosody, sarcasm, hesitation, sighs, emphasis patterns
- Visual signals: facial expressions, gaze direction, head position, micro-gestures
- Temporal signals: how these change over the course of an utterance and the conversation
Output latency for the audio perception pipeline: sub-100ms. Combined pipeline latency (audio + visual + temporal): under 600ms. Context staleness: never exceeds 300ms.
Raven-1 does not output raw classifications. It outputs natural language descriptions: “The speaker sounds surprised and slightly skeptical.” These descriptions feed into the downstream LLM context, allowing the conversational AI to respond not just to what was said but to how it was received.
Tavus acknowledges a limitation that merits attention. In their own published content, they note that Raven-1’s emotion recognition was “trained on English-language datasets” — a limitation that creates “a structural validity problem for global deployments where emotional expression varies by culture.” Cross-domain performance drops to as low as 68.75% in out-of-domain testing. Fear and disgust are the worst-performing emotion categories, likely due to lower representation in training data.
This is not a disqualifying flaw for most use cases. But it is a real limitation for global enterprise deployments or clinical contexts where emotional misclassification has consequences.
The EU AI Act is also relevant here. Restrictions on AI-based emotion detection in workplaces took effect in February 2025. Tavus explicitly calls out this regulatory risk in their own content for hiring applications. Buyers deploying Raven-1 in EU employment contexts should get legal review before going live.
Sparrow-1: Conversational Timing
Released January 2026, Sparrow-1 solves a problem that most conversational AI systems handle badly: knowing when a human is done speaking.
Traditional systems use silence detection: a timer runs, and when silence exceeds a threshold, the AI starts responding. This produces the characteristic awkward pauses of voice interfaces — the system cuts in after a beat of silence, interrupts mid-thought, or fails to respond when someone pauses to think.
Sparrow-1 is an audio-native turn-taking model that operates at 40ms frame-level granularity. It models floor ownership probabilistically — estimating in real time who should be speaking next, based on prosodic signals, completeness of thought, and individual speaker patterns. It adapts to individual speaking styles during the conversation.
Published benchmarks: 55ms median latency, 100% precision, 100% recall, zero interruptions across 28 real-world test samples.
The significance of zero interruptions is harder to appreciate without having experienced conversational AI that interrupts constantly. In high-stakes contexts — a patient asking a medical AI about their diagnosis, a job candidate practicing an interview response, a customer discussing a billing dispute — being interrupted is not just annoying. It is trust-destroying.
Hummingbird-0: Zero-Shot Lip Synchronization
Released April 2025, Hummingbird-0 is the lip-sync layer. Its key differentiator: zero-shot operation, meaning it does not require per-person fine-tuning to synchronize lip movements with audio for a new speaker.
Most alternative lip-sync systems require training a person-specific model before accurate synchronization is possible. Hummingbird-0 achieves state-of-the-art benchmarks without this prerequisite:
| Metric | Hummingbird-0 | Alternative | Direction |
|---|---|---|---|
| FID (Fréchet Inception Distance) | 63.92 | 95.67 | Lower is better |
| LSE-D (Lip Sync Error Distance) | 6.74 | 7.04 | Lower is better |
| ArcFace Identity Score | 0.84 | 0.78 | Higher is better |
The three-stage pipeline: 3D face reconstruction → audio-driven animation → frame synthesis. The zero-shot capability matters most for scale: a platform that requires per-person model training before deployment cannot support large-scale personalized video creation efficiently.
The Integration Thesis
Tavus frames their Conversational Video Interface (CVI) as “replacing a five- or six-vendor stack with a single endpoint.” The competitors being displaced are the configurations that developers previously assembled: a WebRTC provider for video streaming, a STT provider for transcription, an LLM for generation, a TTS provider for voice synthesis, a lip-sync model for avatar animation, and something to tie it together.
Whether this is entirely accurate depends on use case. Enterprise integrations rarely reduce to a single API call. But as developer positioning goes, it is precise — the CVI does integrate all of these layers, and the proprietary models (Phoenix-4, Raven-1, Sparrow-1, Hummingbird-0) provide differentiation that a self-assembled stack from commodity models cannot easily replicate.
The Conversational Video Interface (CVI): What Developers Actually Get
The CVI is Tavus’s primary developer-facing product. It exposes a REST API plus TypeScript, JavaScript, and Python SDKs, with a React component library for browser embedding.
The architecture has two configurable objects:
Replica: the visual entity — the AI avatar. Sourced from a custom replica (created from a 1-minute training video), a stock replica (100+ available on Growth tier), or a photo-based replica.
Persona: the behavioral configuration. Eight configurable elements:
- System prompt (the instruction layer for the underlying LLM)
- Pipeline mode
- Default replica assignment
- STT provider and configuration
- LLM provider and model selection
- TTS provider and voice selection
- Perception layer configuration (Raven-1)
- Conversational flow layer (Sparrow-1)
Plus optional additions: Knowledge Base (RAG, ~30ms retrieval), Objectives (goal-oriented instructions), Guardrails (behavioral boundaries), Memories (cross-session persistence).
Supported LLMs:
- Tavus-hosted: tavus-gpt-oss, tavus-gemini-2.5-flash, tavus-claude-haiku-4.5, tavus-gpt-5.2, tavus-gemini-3-flash
- Custom: any OpenAI-compatible streaming endpoint, including Azure OpenAI
Supported TTS:
- Cartesia (default; sonic-2, sonic-3 models)
- ElevenLabs (eleven_turbo_v2_5 and others)
Supported STT:
- tavus-auto (intelligent routing)
- tavus-parakeet (English/European; lowest latency)
- tavus-soniox (Indian languages)
- tavus-whisper (broad multilingual)
- tavus-deepgram-medical (clinical vocabulary)
Language support: 42+ languages for real-time conversational video.
The daily.co WebRTC layer underlies the real-time video transport. Pipecat integration is documented. LiveKit agent integration is supported. These are the standard developer integration paths for real-time video applications.
Video Generation (Async) and Replicas
Separate from the real-time CVI is Tavus’s async video generation — the capability to produce scripted, pre-recorded videos from a replica. This is closer to what HeyGen and Synthesia do.
Custom replica creation:
- Record a 1-minute training video: approximately 30 seconds speaking, 30 seconds listening-behavior capture
- Photo-based replicas also available (lower fidelity)
- Explicit verifiable consent required for replicas depicting real humans
- Available on Starter, Growth, and Enterprise plans (not the free Basic tier)
Use cases for async generation: personalized video at scale (sales outreach, customer onboarding, L&D content), where real-time interaction is not required but the avatar’s face and voice are used.
PALs: The Consumer Product
In late 2025, Tavus launched PALs — Personal AI companions with persistent memory, emotional intelligence, and real-time video conversation. AI Santa was a holiday demonstration that drove “hours per day” of user engagement per TechCrunch reporting in December 2025.
PALs represent a consumer-facing application of the same CVI infrastructure. The PAL pricing tiers are:
| Tier | Price | Minutes | Notable |
|---|---|---|---|
| Free | $0/mo | 15 min voice/video | — |
| Plus | $20/mo | 150 min | MCP Early Access |
| Max | $50/mo | 500 min | MCP Early Access |
“MCP Early Access” on the Plus and Max tiers is the clearest signal that Tavus is building toward an official MCP server — they are testing it with paying consumer users before general release. As of this writing, no official MCP server is listed in the Anthropic registry or generally available.
Pricing (Developer Tiers)
| Tier | Price | Conversational Video | Video Generation | Custom Replicas | Concurrent Streams |
|---|---|---|---|---|---|
| Basic (Free) | $0/mo | 25 min | 5 min | 3 trainings/mo | 1 |
| Starter | $59/mo + overage | 100 min | 10 min | 3/mo | 3 |
| Growth | $397/mo + overage | 1,250 min | 100 min | 7/mo | 10+ |
| Enterprise | Custom | Custom | Custom | Custom | Custom |
Overage rates:
- Conversational video: $0.32–$0.37/min (varies by tier)
- Video generation: $0.90–$1.00/min
- Custom replica training: $40–$65 per replica
- Conversation recording storage: $0.03/min
Key observations:
The Growth tier at $397/month is the natural production entry point for developers building real applications. 1,250 minutes of conversational video per month at that base rate is sufficient for moderate-volume applications — an interview prep platform serving hundreds of users monthly, a healthcare intake tool, a sales coaching product.
The overage structure creates visibility into marginal cost. At $0.32/min for conversational video, a 10-minute conversation costs approximately $3.20 in overage — meaningful in consumer applications where engagement time is long, more manageable in enterprise contexts where interactions are bounded.
The free Basic tier at 25 conversational minutes is useful for evaluation, not production. It is structured correctly for developers who want to test the integration before committing.
Customers, Use Cases, and Verticals
Tavus does not publish an ARR figure or a total customer count. What they have published are named case studies:
Final Round AI — AI-powered job interview preparation. 100,000+ active users. 1.2 million+ practice interview minutes logged via Tavus. Average session length of 12 minutes. CMO quoted: “Tavus has speed, reliability, flexibility, and the best product on the market.”
This is the clearest validation of Tavus’s real-time thesis in action. Interview practice is a context where the conversational video format is not a novelty — it is the product. You cannot practice job interview responses with a text chatbot in a meaningful way. The video format, the eye contact, the timing, the ability to see your own expression in the frame — these are what make interview prep useful. Tavus is not competing with text here. It is competing with human mock-interviewers.
iAsk — AI tutoring platform. 22,000+ monthly active users of the Tavus-powered video tutor feature. 1.5 million daily queries across the platform. 75% Gen Z user base. Deployed in 1–2 days.
VEED — Video editing platform with 12 million users. Integrated Tavus replicas into VEED’s editor alongside OpenAI, Nvidia, ElevenLabs, and DeepL capabilities. 76% Fortune 500 reach through VEED’s enterprise customer base.
Work Trial AI — AI-powered hiring platform that has generated 10,000+ work trials. Customers include Linear, PostHog, and Automattic. Work Trial uses Tavus for candidate interaction interfaces.
Orum — Series B AI conversation platform for sales coaching.
The pattern across these customers is consistent: real-time video interaction replaces an activity that previously required a human — an interviewer, a tutor, a hiring coordinator, a sales coach. The video format is not decorative. It is functional. The use cases make sense only because the underlying technology can hold a real-time conversation.
Primary verticals: L&D and training, recruiting, healthcare (with HIPAA compliance), sales coaching, customer support, education.
Deloitte and Amazon appear on Tavus’s homepage as enterprise references. These were not confirmed as direct Tavus enterprise contracts in accessible sources — they may be end-customers using Tavus-powered applications from partners. We flag this because Tavus’s own confirmed case studies (Final Round AI, iAsk, VEED) are verifiable, while the enterprise logo presentation on homepages sometimes overstates the directness of the relationship.
Compliance and Security
Tavus holds the following certifications, confirmed via their security trust center:
- SOC 2 Type II
- HIPAA (enabling deployment in clinical and healthcare contexts)
- GDPR compliance
The HIPAA certification is notable for healthcare use cases. A conversational AI intake system at a clinic, a patient education AI that explains diagnoses, or a mental health support companion require HIPAA compliance for U.S. deployments. The tavus-deepgram-medical STT model (clinical vocabulary) paired with HIPAA compliance positions Tavus for healthcare integrations that most AI video platforms cannot support.
Consent requirements: Tavus requires explicit, verifiable consent for replicas depicting real human likenesses. Users are contractually responsible for generated content. This is a reasonable standard for the category, though Tavus has not published a formal Avatar Governance Framework equivalent to Synthesia’s more detailed governance document.
MCP Ecosystem: Close but Not There Yet
The MCP ecosystem question matters for this site’s readers, and the answer is nuanced.
No official Tavus MCP server currently exists in the Anthropic registry or as a generally available product.
What does exist:
-
tavus-skills (github.com/Tavus-Engineering/tavus-skills) — an official Tavus Engineering repo with 8 agent skills installable via
npx skills add. These are Claude Code / agent context skills — they provide instructions and context for agents building with Tavus APIs. MIT licensed. A developer-facing agent enablement tool, not an MCP server in the protocol sense. -
MCP Early Access on PAL consumer tiers — The Plus ($20/mo) and Max ($50/mo) PAL plans list “MCP Early Access” as a feature. This is the clearest signal that an official MCP server is in active development. They are testing it in production with paying users.
-
Community MCP servers — At least two third-party implementations exist on GitHub (one by rakeshdavid, one by xchrismmgx). Both have minimal stars and no official backing. The expected caveats apply: no maintenance guarantees, potential compatibility drift, not supported by Tavus.
Assessment: Tavus is MCP-aware and clearly building toward official MCP support. The PAL tier early access structure suggests it will arrive. For the current moment, if MCP integration is a requirement for your deployment, HeyGen is the better choice — they have a working remote hosted MCP server on all paid plans, handling real-time video generation via tool calls today.
If MCP integration can wait 1–2 quarters, the Tavus early access path may be worth watching.
Tavus vs. HeyGen vs. Synthesia: The Honest Comparison
These three companies are often described as competitors, but they are increasingly solving different problems.
Tavus vs. HeyGen
HeyGen has reached $100M ARR, 31 million registered users, 100,000+ paying businesses, and profitability since Q2 2023. Their Avatar V model is a benchmark leader for async avatar video. Their official MCP server is available on all paid plans. Their 175+ language breadth and video translation feature are best-in-class for async production.
HeyGen’s Interactive Avatar — their closest product to Tavus’s CVI — exists and is deployed. But HeyGen’s core business, heritage, and infrastructure are built for async video creation. Interactive Avatar is a feature built on top of that foundation.
Tavus’s core business, heritage, and entire model stack are built for real-time conversational video. CVI is the foundation, not a feature on top.
HeyGen wins when: you need async video at scale, a large template library, 175+ language support, a consumer-friendly interface, an official working MCP server today, or a product priced for individual creators.
Tavus wins when: you are building a real-time application where the conversation itself is the product — interview prep, AI tutors, customer service agents, sales coaching, medical intake — and you need sub-500ms latency, emotional perception, and turn-taking that feels natural.
Tavus vs. Synthesia
Synthesia is the enterprise async video standard. $100M ARR, ISO 42001 certification (first in the category), SCORM export for LMS compliance training, 50,000+ customers, 90%+ Fortune 100 penetration. Synthesia’s product is about producing training videos at scale, not having conversations.
Synthesia does not offer a real-time conversational product. Their “Video Agents” feature was listed as “coming soon” in research conducted for this site. Synthesia’s governance apparatus — Avatar Governance Framework, formal consent recording, 24/7 Trust & Safety team — is more developed than Tavus’s published documentation.
Synthesia wins when: you are in enterprise L&D, require SCORM export for LMS compliance, need ISO 42001 certification, or are deploying across a Fortune 500 organization that has already standardized on Synthesia.
Tavus wins when: you need a real conversation, not a video playback. The use cases do not overlap much.
Where the Category Is Going
The emerging architecture for enterprise AI is not “choose async video OR conversational video.” It is “use the right tool for each touchpoint.” A compliance training module where the content needs to be identical for every employee — Synthesia. A sales coaching practice session where the AI needs to respond to what the rep actually says — Tavus. A viral creator-market video of a CEO announcing a product — HeyGen.
The risk for HeyGen is that their Interactive Avatar feature faces more direct competition from a Tavus that is getting better funded and more capable. The risk for Synthesia is that “coming soon” Video Agents is playing catch-up to a company that has spent four years building the real-time stack they would need to replicate.
Limitations and Risks
No publicly disclosed revenue or scale metrics. Tavus does not publish ARR, customer count, or user count. For a Series B company, this is notable. The most charitable interpretation: developer infrastructure companies sometimes defer public metrics until they have a cleaner story to tell. The less charitable interpretation: the numbers are not yet impressive relative to HeyGen’s $100M ARR. Buyers should treat Tavus as a technology bet with strong technical foundation but uncertain commercial scale.
No official MCP server yet. The gap is acknowledged and actively closing, but it is a gap. Developers building agent-native applications today need to use community integrations or wait for Early Access to expand.
Emotional perception limitations. Raven-1’s English-language training data bias, cross-domain performance dropping to 68.75%, and EU AI Act restrictions on emotion detection in workplaces are real constraints for global or regulated deployments.
Small team relative to ambition. ~40 employees is a small team for the infrastructure layer it is trying to build. If Phoenix-4 requires significant ongoing research maintenance while CVI enterprise support demands scale, the headcount may create bottlenecks.
Deepfake and misuse risk. Tavus operates in face-and-voice cloning. Their consent requirements are reasonable; their governance documentation is less developed than Synthesia’s. This is a category-level risk rather than a Tavus-specific failure, but buyers in regulated industries should review the terms and compliance posture carefully.
PAL consumer product: Adding a consumer AI companion product (PALs) to a developer infrastructure business is a strategic choice that comes with questions about focus. The engagement numbers from AI Santa are encouraging; whether the PAL consumer business competes productively with or distracts from the developer CVI platform is worth watching.
What to Build With Tavus
The developer use cases where Tavus is distinctively suited:
AI interview coaching: Practice job interviews with a real-time AI interviewer who maintains eye contact, listens to your answers, responds to what you actually said, and uses Sparrow-1’s turn-taking so it does not cut you off mid-sentence. Final Round AI has demonstrated this works at scale.
AI tutors: The iAsk deployment — 22,000+ monthly active users on a video tutor — validates the education use case. Subjects that benefit from dialogue (language learning, medical education, professional certifications) map well to conversational video.
Healthcare intake and patient education: HIPAA certification + tavus-deepgram-medical + Raven-1 emotional perception makes Tavus the most complete stack for healthcare conversational video applications. A patient intake AI that can perceive when someone sounds anxious, a symptom-collection agent that holds a natural conversation, or a post-procedure care instructions AI are all deployable today.
Sales coaching and roleplay: Simulating a customer objection conversation in real time, with a realistic AI who gives pushback and reads tone, is a training use case that text cannot replicate. Orum has deployed Tavus in exactly this context.
Customer service agents with a human presence: Voice-only or text chat customer service is already broadly deployed. Video-first customer service — a face, eye contact, emotional responsiveness — is an emerging category. Tavus is the best-positioned infrastructure for developers building in this space.
The Verdict
Tavus is doing something genuinely hard. Real-time emotionally intelligent conversational video AI with sub-500ms latency, turn-taking that does not interrupt, and a renderer that generates micro-expressions per-frame is not assembled from off-the-shelf components. The model stack — Phoenix-4, Raven-1, Sparrow-1, Hummingbird-0 — represents four years of focused technical development, each with published benchmarks, each solving a specific subproblem that matters in production.
The funded developer infrastructure play is credible. Sequoia does not typically back companies without commercial traction. The named customers — Final Round AI at 1.2M practice minutes, iAsk at 22,000 monthly users, VEED at 12M users integrating Tavus — are genuine production deployments, not beta logos.
The limitations are also real. No ARR disclosure makes the commercial scale difficult to assess. No official MCP server is a current gap. Emotional perception has documented limitations that matter for global and regulated deployments. The ~40-person team is lean for an infrastructure ambition.
The rating is 4 out of 5. A full score would require either publicly demonstrated commercial scale comparable to HeyGen and Synthesia, or a mature MCP server and governance documentation at Synthesia’s level. What earns the 4 is genuine technical differentiation, clear category focus, and a thesis that the use cases validate: real-time conversational video AI is a distinct category, it is larger than it appears today, and Tavus has built the most capable technical foundation for developers who want to build in it.
If you are building an application where the conversation itself is the product — not a video you produce but an AI you talk to — Tavus is where you start.
ChatForest researches AI tools from public sources, documentation, and published benchmarks. We do not test tools hands-on or make claims about direct product experience. This review reflects information available as of May 2026.