Tavus

Most AI avatar tools turn a script into a video. Tavus does something harder: it puts a face on an AI agent and lets it hold a live video conversation. The agent sees the person through their camera, hears tone and hesitation, and answers in real time.

Tavus calls these agents PALs and describes itself as an AI research lab working on "human computing." Its newest model, Griffin, is the headline on its homepage. Tavus says 48% of people in its test thought they were talking to a real person, and calls it the first model to pass a real-time video Turing test. That's the company's own study. It does show where the product is aiming.

Plans priced in conversation minutes

Monthly prices on October 6, 2026. Business, at $975, and Enterprise appear further right on the page.

Every plan runs at 1080p with 24 kHz audio and covers more than 40 languages. Paid plans remove the watermark. Growth and Business add conversation recordings at $0.03 a minute, and Enterprise brings volume discounts, custom concurrency and a dedicated Slack channel.

What a real deployment costs

The included minutes get cheaper as plans grow: about $0.37 a minute on Starter, $0.34 on Builder, $0.31 on Growth and $0.24 on Business. The jump between tiers matters more than the per-minute rate, though.

Take a sales agent that talks to visitors for 30 hours a month, or 1,800 minutes. On Growth, that's $397 plus 500 extra minutes at $0.31, about $552. Business would cost $975 and leave 2,200 minutes unused. Business makes sense from roughly 3,200 minutes a month, where Growth's overage would push past it.

Starter is a sandbox with a hard ceiling. With no pay-as-you-go option, conversations simply stop at 60 minutes a month, and each one ends after five minutes. That's enough for a demo, not for anything customer-facing.

What a PAL can do

A face and a voice are only the start. Tavus's models handle the parts that make a conversation feel natural. Phoenix-4.5 generates the full face with emotion control. Raven-1 reads expressions and surroundings, and Sparrow-2 decides when to speak and when to wait.

On top of that sit practical skills. A PAL can present a slide deck on its own, generate charts or forms on screen with what Tavus calls Magic Canvas, share its screen while browsing your product, and join Google Meet, Zoom or Teams calls as a participant. It answers from a knowledge base you upload, remembers people between conversations and can call your own functions and tools. You can use Tavus's hosted language models or bring your own.

Developers build through the API. Everyone else can use the PAL Maker, which creates a working agent from a description without code. Tavus says more than 150,000 developers use the platform, and it cites Salesforce, which built a sales-development agent on it.

Recorded videos too

Tavus started with personalized recorded videos, and that product is still included. Paid plans come with 5 to 250 minutes of generated video a month. Extra minutes cost $1 on Builder, $0.90 on Growth and $0.80 on Business.

For recorded avatar videos alone, specialist tools cost less. HeyGen starts at $24 a month billed yearly and Synthesia at $18, and both offer large libraries of presenters. D-ID also sells talking agents, from $4.70 a month on yearly billing. The AI avatar generators guide works out what a finished minute costs in each.

Whichever plan you choose, watch usage closely in the first weeks. A live agent that people enjoy talking to can use up a month's minutes faster than any forecast.