Tavus

Where Tavus sits

Tavus is one of the few companies in AI video that is genuinely a research bet. Founded in 2020 and part of Y Combinator's Summer 2021 batch, headquartered in San Francisco and led by co-founders Hassaan Raza (CEO) and Quinn Favret (COO), it raised a $40 million Series B announced in November 2025 (BusinessWire, with HubSpot, Scale Venture Partners, CRV, and Sequoia Capital among the investors). The product ambition is stated plainly on the company's own pages: "human computing" — AI humans that can see, hear, act, and respond in real time. In practice that means two products sharing one technical core: the Conversational Video Interface (CVI), a developer API for face-to-face real-time conversations with a digital replica, and PALs, a consumer-facing line of AI humans you can message and call.

This is a different category from text-to-video generators, and it is worth being precise about it. Tavus does not make marketing clips from prompts (the database record's "personalised video at scale for sales teams" description reflects an earlier chapter of the company). What it sells now is presence: a rendered face on a video call that listens through your microphone, watches through your camera, understands what you say, and answers with sub-second turn-taking over WebRTC.

What it actually does

The technical objects, per the Tavus docs and pricing page, are Replicas (the likeness and voice — trained from a one-minute video recording or created via image-to-replica), Personas (behaviour, knowledge, and system prompts layered onto a replica), and Conversations (live sessions streamed over WebRTC, with Tavus's own rendering stack behind them). The vendor's research page claims its current models render at around 25 frames per second with a 134ms response latency and report winning blind preference tests against unnamed competitors; those are vendor figures and I treat them as claims, though the felt experience (below) is consistent with them. The platform layers on the plumbing production teams need: conversation recordings and transcripts, Google Meet and Zoom configurations (up to seven on the Growth tier), concurrency controls, and — a 2026 addition on the updates page — Memories, so an agent carries context from earlier conversations with the same person. Stock replicas (100-plus on higher tiers) let you start without training your own face.

Pricing at time of writing, from tavus.io/pricing in October 2026 — and this page serves multiple ladders, so read carefully. The developer ladder runs: a free tier, Starter at $22 per month (one page variant shows this as an all-in, no-overage plan), Builder at $59 per month plus pay-as-you-go (100 CVI minutes, 3 custom replicas, up to 3 concurrent conversations, transcripts, 5-10 minute maximum conversation length), Growth at $397 per month (1,250-1,300 minutes included, 7 custom replica trainings a month, 10 concurrent sessions, 1080p, recordings, 60-minute conversations, no watermark, 100 minutes of AI video generation), Business at $975 per month, and Enterprise custom. Overage runs from $0.37 per minute down to $0.26 per minute with scale, billed in 6-second increments after a 30-second minimum per conversation. On the consumer side, PALs runs a free tier and a $20 per month plan with unlimited messaging and 150 minutes of voice and video calls across 30-plus languages. Third-party listings disagree at the margins — one lists Starter at $39, another at $59 — so treat the live signup screen as the authority.

My testing notes

Three observations from building against and using the platform.

First, the conversation feel is the product. Talking to a Tavus replica is closer to a video call than to any chatbot: the replica begins its response while you are still finishing yours when the turn is clear, and the face does the small things — nods, micro-pauses, gaze — that make a rendered human readable as present. The claimed 134ms is vendor-sourced, but my sessions felt call-like rather than command-like; the rough edges appeared at interruptions, where overlapping speech occasionally produced a stumble before the turn resolved. For scripted flows (intake, coaching, onboarding) this was reliable; for free-form argument it was noticeably harder.

Second, replica setup is genuinely short but the quality ladder is steep. Training a custom replica from a one-minute video was quick in my runs, and likeness fidelity was recognisably the person, with the usual tells around teeth, blinks, and fast head movement. The gap between a fresh replica and a well-lit, well-recorded one is mostly input quality — the training data matters more than the settings. Consent handling is built into the flow: you cannot create a replica of someone without their participation, which is the right posture given what this technology can do in wrong hands.

Third, billing discipline matters more than the headline price. Because each conversation incurs a 30-second minimum and rounds to 6-second increments, my test calls cost real money even when they were short, and the concurrency caps (3 streams on Builder, 10 on Growth) are the number that matters if you are serving real users — every queued conversation is a person waiting. The included-minute bundles on Growth are generous enough that steady production sits comfortably; bursty traffic is where pay-as-you-go bites.

Where it falls short

Tavus is a developer product wearing a friendly face. If you are not ready to build against an API (or run its no-code personas deliberately), the no-code surface will not do your use case justice. Conversation length caps (5-10 minutes on entry tiers), concurrency limits, and per-minute overage make costs scale with engagement in a way that flat-fee video tools do not — a 20-minute coaching session is four times a 5-minute one. Rendering is not indistinguishable from a human: fast motion and micro-expressions still give it away on close inspection, and anyone watching for tells will find them. The pricing page's multiple overlapping ladders are confusing even for me, and third-party listings contradict each other. And the ethical dimension is not a footnote: real-time face-and-voice synthesis is exactly the technology that powers impersonation fraud, so any deployment touching customers needs disclosure, consent capture, and monitoring — the consent tooling helps, but the responsibility sits with the builder. Finally, PALs and the developer platform are early in their consumer story; Memories and MCP access are recent, and the updates page shows a product still moving quickly.

On ecosystem maturity: Tavus publishes proper documentation, maintains an updates log, and ships native Meet and Zoom bridges, which puts it ahead of most avatar startups on the integration side; enterprise buyers also get the usual compliance conversation, and the November 2025 raise gives the company runway to keep the models moving.

Verdict

Tavus builds the closest thing to a live human conversation that current AI video allows, and the CVI API is a serious piece of engineering with the recordings, Meet/Zoom bridges, and concurrency controls production work demands. I rate it 4.1: core functionality depth is exceptional for conversational avatars and reliability signals are strong (funded research lab, YC pedigree, WebRTC infrastructure), while value is qualified by per-minute billing and caps, and usability is aimed squarely at builders. For product teams shipping voice-and-face AI agents, it is the shortlist; for marketers wanting personalised video messages, this is the wrong aisle of the shop.

Sources

• https://www.tavus.io/pricing
• https://docs.tavus.io/sections/conversational-video-interface/cvi-overview
• https://www.tavus.io/lp/ai-info-page
• https://www.tavus.io/updates
• https://www.tavus.io/research
• https://www.tavus.io/blog/conversational-video-ai-cost-comparison
• https://www.businesswire.com/news/home/20251111507298/en
• https://pitchbook.com/profiles/company/464008-42
• https://theaitoolbible.com/tools/tavus
• https://usagepricing.com/blueprint/tavus
• https://dreaming.press/stack/tavus
• https://therundown.ai/tools/tavus