Tavus AI: How to Use It, Pricing, and What the New Griffin Model Means (2026)

Tavus AI makes real-time video AI: digital humans that can hold a live face-to-face conversation, see and hear you, and reply with a lifelike face and voice. The company is in the news after announcing Griffin on 1 October 2026, a model that Tavus says fooled nearly half of the people in its own test into thinking they were talking to a real person. This guide explains what Tavus AI is, how to use it today, what it costs, how it compares with alternatives such as HeyGen and Synthesia, and what Griffin does and does not change.

Everything below was checked on 3 October 2026 against Tavus’s official pricing page, its Griffin announcement page and the Tavus developer documentation. Figures about Griffin’s performance come from Tavus itself and have not been independently verified.

What Is Tavus AI?

Tavus is a San Francisco company building what it calls “human computing”: AI you talk to the way you would talk to a person on a video call. Instead of typing into a chatbot, you speak to an on-screen face that listens, watches your expressions through the camera and responds in real time.

The platform has two sides:

• PALs for consumers. PAL stands for Personalized AI Likeness. These are ready-made AI companions and assistants you can message, call or video chat with in more than 30 languages.
• The developer platform. Businesses use Tavus’s API to put their own video agents into websites and apps, such as an AI sales rep, a tutor, an interviewer or a customer-support agent.

The Tavus models

Tavus builds its own family of models, each handling one part of the conversation:

• Phoenix-4.5 renders the face. It creates a photorealistic “replica” (Tavus now also calls these Faces) from a short training video or even a photo.
• Raven-1 is the perception model. It understands what it sees and hears during the call, so the agent can react to your expression or what is on screen.
• Sparrow-2 handles conversational flow and turn-taking, deciding when to speak and when to listen.

Tavus’s pricing page says all paid developer plans include Phoenix-4.5, Raven-1 and Sparrow-2.

What Is Tavus Griffin?

Griffin is the headline of the week. Tavus describes it as “the first Human Interaction Model”, a single model that processes audio and video at the same time instead of passing your words through separate speech-recognition, language and rendering steps. In theory, that lets it interrupt, nod, react and respond the way people do, rather than waiting for a clean turn.

The 48% “video Turing test”

According to Tavus, 54 participants were told they would video-call another person. Their partner was actually Griffin-Lite. After a one-minute call, 48% (26 of 54) said they believed it was a real person. By comparison, Tavus says only 2.4% believed this of its previous system. Participants who did suspect AI usually decided within about 20 seconds.

Treat those numbers with care. It was a small study, run by Tavus itself, with one-minute calls and no independent evaluator. It is an impressive claim, not settled science.

Can you use Griffin yet?

No. Tavus says Griffin-Lite is a research preview available only to select trusted testers, and that it “will not be available for use for customers at this time”. There is no public API, price or release date. Tavus’s reason is safety: a system convincing enough to pass as human could also be used to deceive people, so the company says it is working on disclosure features before a wider release.

In practice, if you want to use Tavus AI today, you will use the current Phoenix, Raven and Sparrow models described above.

Tavus Pricing (Checked 3 October 2026)

Tavus pricing is split between consumer PAL plans and developer plans. All prices are in US dollars.

PALs (consumer) plans

Developer plans

Other differences between developer plans:

• Custom replicas: none on Free, 3 trainings a month on Starter, 7 on Growth, unlimited on Enterprise.
• Stock replicas: 25 on Free and Starter, 100+ on Growth.
• Enterprise extras: SOC 2 and HIPAA compliance options, SLAs and dedicated support.
• Conversation recording: listed at $0.03 a minute on Growth.

Is Tavus free? Yes, up to a point. Both the consumer and developer sides have free plans, which are enough to test it. Real deployments quickly need a paid plan, because live conversation is billed by the minute. A customer-facing agent that talks for 5,000 minutes a month would exceed the Growth allowance, so check your expected volume against the overage rates before choosing a plan.

How to Use Tavus AI: Two Routes

Route 1: Talk to a PAL (no code)

1. Go to tavus.io and sign up for the free PALs plan.
2. Choose a PAL and start with text messaging to get a feel for its personality.
3. Start a voice or video call. The free plan includes 15 minutes of calls.
4. Upgrade to Plus or Max only if you use it regularly.

This is the quickest way to see what real-time conversational video feels like before building anything.

Route 2: Build your own video agent (developers)

Tavus’s documentation describes a simple flow using its REST API at tavusapi.com:

1. Get an API key from the developer dashboard. Every request sends it in an x-api-key header.
2. Create a PAL with POST /v2/pals. Give it a name, a system prompt describing how it should behave, and a face. The docs use the stock “Anna” face (ID rc9cff32ceba) so you can start without training anything.
3. Configure the layers if needed: perception, speech-to-text, the language model, text-to-speech and conversational flow. The defaults are designed for low latency.
4. Start a conversation with POST /v2/conversations, passing your PAL’s ID.
5. Join the call by opening the conversation_url in the response, or embed it in your own site or app.

If you already use LiveKit or Pipecat for voice agents, Tavus documents integrations for both, so you can add a Tavus face to an existing pipeline.

Creating a custom replica of yourself

On paid plans you can train a custom face from a video of yourself or, with Phoenix-4.5, even from a photo. Tavus’s docs say a zero-shot preview is usable within about a minute from a photo or a few minutes from video, carrying a visible watermark while the model finishes tuning in the background. Phoenix-4 is still offered for full-body replicas.

You must have the rights to every face and voice you upload. Tavus’s documentation is explicit that you need the necessary permissions for any likeness you use. In practice, only clone yourself, or someone who has clearly agreed in writing.

What Can You Use Tavus For?

• Customer support and sales: a video agent that answers product questions on your website around the clock.
• Training and education: a tutor or role-play partner for language practice, interview preparation or sales training.
• Recruiting: structured first-round screening interviews.
• Personalised video at scale: Tavus’s video-generation minutes create recorded videos with your replica, useful for outreach and onboarding.
• Healthcare intake and wellbeing apps: with Enterprise compliance options.

For a wider look at how businesses are deploying agents like these, see our guide to how AI agents are changing business in 2026.

Tavus vs HeyGen, Synthesia and Other Alternatives

“Tavus vs HeyGen” and “Tavus alternatives” are common searches, but the tools are not all doing the same job.

The simplest rule: if you need a live conversation, look at Tavus; if you need a finished video file, look at an avatar video generator. Our overview of AI avatar video generators covers the recorded-video side, and our report on Gemini Live Avatar explains Google’s competing approach. If voice quality is your main concern, our guide on how to use ElevenLabs covers voice cloning in depth.

Is Tavus Safe? Privacy and Ethics

Griffin has pushed a real question into the open: if a video AI can pass as human, how will people know who they are talking to? A few practical rules for anyone using Tavus or similar tools:

• Always disclose. Tell people at the start of a call that they are speaking to an AI. In the EU, the AI Act includes transparency obligations for systems that interact with people and for deepfake content, so disclosure is becoming a legal matter as well as an ethical one.
• Get consent for every likeness. Never create a replica of someone without their explicit permission.
• Think about recordings. Video calls capture faces and voices, which is personal data under GDPR. Check where conversations are stored, for how long, and whether recording is switched on.
• Keep humans in the loop for sensitive decisions such as hiring, healthcare or finance.

Tavus’s own decision to keep Griffin in a restricted preview suggests the company takes these risks seriously. It also shows the technology is moving faster than the safeguards around it.

Tips for Better Results With Tavus AI

• Write a focused system prompt. One clear role (“a friendly onboarding guide for our booking app”) works better than a vague all-rounder.
• Start with a stock face while you design the conversation, then train a custom replica once the script works.
• Record good training footage: good light, a steady camera, a plain background and natural speech.
• Keep conversations short while testing, especially on the free plan’s 5-minute limit.
• Track minutes from day one so overage costs do not surprise you.

If you plan to turn Tavus recordings into marketing content, our guide to AI video marketing covers how to use AI video without losing your audience’s trust.

FAQ

What is Tavus AI used for?

Tavus AI is used to build real-time video agents and AI companions that hold face-to-face conversations, plus personalised recorded videos made with a digital replica.

Is Tavus AI free?

There are free plans. Consumers get 15 minutes of calls a month, and developers get 25 minutes of conversational video and 5 minutes of video generation. Paid developer plans start at $59 a month.

Can I use Tavus Griffin?

Not yet. Griffin-Lite is a research preview for selected testers only, and Tavus says it is not available to customers while it works on safety and disclosure features.

Is Tavus better than HeyGen?

They solve different problems. Tavus focuses on live, two-way video conversations, while HeyGen is mainly used to create recorded avatar videos. Choose based on whether you need a conversation or a video file.

Does Tavus work in Europe?

Tavus is a US company with prices in US dollars, and PALs support more than 30 languages. Businesses in the UK and EU should review data handling, recording settings and AI-disclosure rules before deploying it with customers.

Final Thoughts

Tavus AI is one of the most advanced platforms for real-time conversational video, and Griffin shows where it is heading: digital humans that are hard to tell apart from people. You cannot use Griffin yet, but you can try Tavus today for free, either by calling a PAL or by building a simple video agent with the API. Start small, disclose clearly that users are talking to AI, and model your minute costs before you scale.

Sources: Tavus pricing; Tavus Griffin announcement; Tavus CVI quickstart; Tavus replica documentation; CellCog analysis of the Griffin study. Checked 3 October 2026.