# Tavus Griffin (HIM) Case Study: The Video Turing Test Launch | Doomers AI

Source: https://doomers.ai/work/tavus-griffin-case-study
Updated: 2026-11-10

> Tavus Griffin case study: how Doomers AI amplified the launch of a model 48% of people mistook for a human, from 11,500 followers to 20M+ views, 30,000 organic posts and a company account that tripled in a day.

# Tavus Griffin: amplifying the launch of a model that *passes for a person*

**21.3M** Views on the post

**44K** Followers, from 11,500

**30,000+** Organic posts about Griffin

**#36** Of 2512 launches tracked

**On 1 October 2026 Tavus introduced Griffin, a model it says passed the video Turing test, from a company account with about 11,500 followers.**

The post stands at 21.3M views, ranked 36 of the 2512 launches we track, and the account that posted it has 43,586 followers, three times what it had the morning before. Tavus ran the launch with partners it brought in.

Doomers AI handled the amplification: more than 150 founders, creators and builders matched to the brief, developers, AI model accounts and the tech generalists everyone follows among them, each timed in how they engaged, so the post met the timeline from every direction at once. Within a day the timeline had answered with more than 30,000 organic posts about Griffin.

This is how that was done, what the post did hour by hour through its first day, and where it landed against every AI launch on X this year. The figures follow the launch directory and update every half hour while the launch is moving.

By [Doomers AI](https://www.linkedin.com/company/doomers-ai) Launch distribution for AI companies

Case study · launched 1 October 2026 · figures follow the launch directory

## The Tavus Griffin launch

At 16:59 UTC on 1 October 2026 the Tavus company account posted four sentences and a video. Griffin is a full-duplex video-to-video model: it listens, watches and speaks at the same time, and generates the whole frame as it goes. Tavus calls it the first Human Interaction Model.

The first sentence is the claim. The first model to pass the video Turing test. In a blind study of 54 people, each on a one-minute video call, 48% said afterwards that they had spoken to a real person.

Earlier systems, Tavus says, passed at most 2% of the time. On NVIDIA's VideoFDB benchmark for full-duplex video, Griffin scored 3.83 out of 5 against a human reference of 3.92, with the next published system at 2.80.

Nobody can use it. Griffin is a research preview, and the last post in the thread says why: a model that can be mistaken for a person needs safety work before a public release. The link goes to the preview that sets out how it was built and tested.

Tavus has raised about $64M: a $6M seed led by Sequoia in 2023, an $18M Series A led by Scale Venture Partners, and a $40M Series B led by CRV in November 2025 with Sequoia, Scale and Y Combinator taking part.

It came through Y Combinator. It had launched before from the same account. Phoenix-4.5, its rendering model, went out on 10 September and finished at 182K views over three weeks.

**Launching a model?** We will read your plan against every model launch in the directory and tell you what the ones above the line led with.

[Plan your launch](https://doomers.ai/contact)

## What Griffin is, in Tavus's words

Tavus calls itself the human computing company. Its argument is that computing has always made people learn the machine's interface, and that the next interface is a face-to-face conversation: software that adapts to a person's language, gestures and emotions in real time.

Its stack until now was three models behind a Conversational Video Interface, or CVI: Raven-1 for perception, Sparrow-2 for conversational timing, and Phoenix-4.5 for rendering a face.

Griffin folds those jobs into one. Tavus defines a **Human Interaction Model (HIM)** as "a new class of model designed to understand and generate face-to-face real-time human interaction," and Griffin is the first: a **full-duplex video-to-video model** in which perception, deciding when and how to respond, and generating speech and video all happen at the same time. It listens while you talk, reads expressions and pauses, and generates the whole frame as it goes.

**The video Turing test.** In a live study, 54 participants each had a one-minute video call with Griffin without being told what it was. 48% believed afterwards that they had spoken to a real person.

Tavus says earlier systems, including its own CVI stack, had a maximum pass rate of 2%. The test it describes asks a different question from the classical one: whether the AI feels human, which includes building rapport, showing empathy and taking initiative.

**The benchmark.** On NVIDIA's VideoFDB benchmark for full-duplex video, Griffin scored 3.83 on generation against a human reference of 3.92 and 2.80 for the next-best system, and 3.73 on perception against 4.20 for humans and 3.44 for the strongest baseline.

On takeover-rate alignment, how closely the model's decisions about when to speak match a person's, it scored 62.8% on generation and 73.8% on perception, the highest of any system.

**Availability.** What shipped on launch day is Griffin-Lite, a research preview open to a select group of early testers, with a wider release of a more powerful model to follow.

Tavus's reason for holding the full model is stated plainly on its page: the same properties that make a Human Interaction Model a strong interface "allow them to deceive a human into believing it is not AI," and further alignment and safety work is required before a safe release. The full research page is at [tavus.io/griffin](https://www.tavus.io/griffin).

That is the product the brief had to launch: a new category with a name nobody had searched for the day before, a claim that invites argument, and a preview most people could not try. Each of those shaped who the post had to reach.

## The brief: who Griffin had to reach

A launch post from an 11,500-follower account reaches 11,500 people on its own. Everything past that is decided in the first hour by who engages, how, and in what order, because the ranking model reads the early response and decides how far to send the post.

The goal was timeline dominance: for the hours after the post, anyone in tech opening X should meet Griffin from more than one direction. A developer sees it from a builder they follow. A founder sees it from another founder with a take.

Someone who follows the big generalist accounts in tech sees it there. The same person sees it twice, from different people, and the ranking model reads that as the thing everyone is talking about, because it is.

The claim set the angles. "The first model to pass the video Turing test" is specific, falsifiable and attached to a number, which is a claim people argue with, and the demos in the replies settle the argument for anyone who watches them. So the brief attacked from every angle, with the weight on four:

- **Developers** who build conversational and video products, and would read the thread as something to build on.

- **Founders and operators in AI** whose take on the Turing test claim is what their followers read first.

- **AI model and research accounts**, the people who cover every model launch and whose audience judges a benchmark claim on sight.

- **Tech generalists with large followings**, the well-known accounts the whole space follows, who carry a launch to people outside the category.

Each angle reaches a different slice of the same timeline, and the point is the overlap. Reach from one direction is a post people notice. Reach from four at once is a day people remember.

## How Doomers AI amplified the Tavus launch

### Matching people to the brief

More than 150 founders, creators and builders were matched to the brief. Each angle was filled from the network on two tests: whether an account's audience is the slice of the timeline that angle is for, and whether its own posting history shows it engages with launches like this one.

The second test matters as much as the first. A founder who has quoted the last three model launches with an opinion will quote this one, and their followers expect it.

A generalist who covers what the whole space is watching will cover this, because the whole space will be watching. An account that never does either would look staged, and the ranking model is good at staged. [Creator amplification](https://doomers.ai/guides/creator-amplification) covers how the match is made.

### Timing the engagement

The second decision is what each person does, and when. A quote post with a take argues with the claim, and the argument is what makes a launch travel.

Quote posts with a take went first, inside the first hour, while the Turing test claim was fresh and the thread was still being read as a question.

Quote posts from accounts with overlapping audiences followed as each demo landed in the replies, so the thread kept returning to feeds through the afternoon. By the time the organic pickup arrived, there was already a conversation to join.

### What we did not control

The claim, the demos and the safety hold were Tavus's. The organic pickup past the 150 seeded accounts was the market's: more than 30,000 organic posts about Griffin across the timeline in the first day, from people nobody briefed.

The amplification put the post in front of the whole timeline, from several directions, in the hours it mattered. The table below is what the timeline did with it.

## The first day of the Griffin post, hour by hour

The directory read the post every half hour from the moment it went in. These are the readings, from the X API, with the time since the post.

| Time after post | Views | Quote posts | Likes | Bookmarks | Followers |
|---|---|---|---|---|---|
| 32 min | 171K | 196 | 536 | 256 | 11,428 |
| 55 min | 368K | 355 | 1,242 | 624 | 11,705 |
| 1 h 44 | 674K | 556 | 3,094 | 1,638 | 12,437 |
| 2 h 35 | 1.17M | 838 | 5,197 | 2,990 | 13,613 |
| 3 h 40 | 2.69M | 1,268 | 9,875 | 6,320 | 16,672 |
| 3 h 59 | 3.23M | 1,430 | 11,307 | 7,366 | 17,634 |
| 6 h 01 | 6.05M | 2,132 | 17,277 | 11,796 | 21,945 |
| 8 h 04 | 7.69M | 2,585 | 20,655 | 14,313 | 24,814 |
| 10 h 16 | 9.09M | 2,904 | 23,330 | 16,158 | 27,099 |
| 13 h 44 | 11.1M | 3,450 | 28,045 | 19,447 | 31,062 |
| 16 h 57 | 12.6M | 3,837 | 30,774 | 21,415 | 33,703 |
| 18 h 52 | **13.3M** | **4,046** | 31,854 | 22,219 | **34,790** |

Four things in that table are the launch.

**The quote posts came first.** At 32 minutes the post had 196 quote posts against 536 likes, more than one quote for every three likes.

The directory median is one for every twenty. That ratio is the first hour of the amplification: people with a take, in public, while the claim was still a question.

**It sped up instead of slowing down.** Most launches do most of their views in the first hour and taper. Griffin did 171K in its first half hour, 500K in its second hour, and 1.5M between hours three and four.

Half a million views arrived in the nineteen minutes between the fifth and sixth readings. That is the shape of organic pickup compounding on a conversation that was already running, and it is the shape the first hour was built to produce.

**It held that pace through the night.** The post added 2.8M views between hours four and six, another 3M by hour ten, and passed ten million while the US was asleep.

At the nineteen-hour mark it was still adding about a million views every two hours, an average of 700,000 an hour for most of a day. A launch that peaks in its first hour and tapers is the median shape in the directory. This one ran flat out.

**The account tripled.** Tavus went from about 11,500 followers to 34,790 in nineteen hours, more than 23,000 people who wanted the next post, and the follower line rose with the views the whole way at roughly 1,750 new followers per million views.

For a company that cannot let anyone use the model yet, that list is what the launch was for.

## Tavus Griffin by the numbers

| Measure | Griffin |
|---|---|
| Views on the launch post | **21,265,868** |
| Quote posts | **5,643** |
| Likes | 39,121 |
| Replies | 3,437 |
| Reposts | 4,269 |
| Bookmarks | 26,913 |
| Reach against the account's following | **1,849 times** |
| Rank among the launches we track | 36 of 2512 |
| Saves per 10,000 views | 12.7 |
| Reposts per million views | 201 |

Read the last two rows against the directory baselines in [the Q3 report](https://doomers.ai/research/state-of-ai-launches-2026#engagement): a median of 12 saves per 10,000 views and 251 reposts per million.

Read the whole table against Phoenix-4.5, the same account three weeks earlier, which finished at 182K views with 61 quote posts. The reach row is measured against the account's following today. Against the roughly 11,500 followers it had when it posted, the figure is over 1,100 times.

## Where the Tavus launch ranks among 2026 AI launches

The [launch directory](https://doomers.ai/launches) holds every AI launch on X we can find this year, 2512 of them, read from the X API. On 2 October, its second day, the Griffin post ranked **36**, and the readings below are from that morning.

- **Most of the launches above it came from accounts with more than 100,000 followers.** OpenAI, SpaceX and Anthropic's Claude account are among them. Griffin is the largest launch in the directory from a company account with under 50,000 followers, and one of 25 posts all year to pass ten million views.

- **On its second day, only three launches had been bookmarked more**, among them OpenAI's 3 September post and TypeSafe AI's. 26,913 people saved a post about a model they cannot use yet.

- **5,643 quote posts** against a directory median of 34, the seventh most quoted launch of the year. Quotes are the measure of argument, and argument is what the first hour was built to start.

- **About 50 times the directory median** of 259K views, and 73 times Phoenix-4.5, the same account's launch three weeks earlier, which sits at 462 of 2512.

Same account, same company, three weeks apart, a 73-fold difference. The product deserved it, and the amplification made sure the timeline found out from every direction in the hours that decide how far a post travels.

## The Griffin thread, demo by demo

The post is four sentences and a 1:49 video: the claim, the number, the benchmark with NVIDIA's name on it, and the category. Then six demos and one explanation, each a reply to the post above it, each filmed live with a real person:

1. Two people try to trip Griffin up at Simon Says. It has to listen for the phrase, watch their hands and move its own, at the same time.

2. While someone is still telling it about a boring date, it laughs, nods and asks questions. Full duplex, on camera.

3. Mid-story, a peacock is held up, and Griffin writes it into the plot. What it sees changes what it says.

4. It coaches a Rubik's cube solve from what it sees in the hands, and waits when the solver goes quiet to think.

5. It talks someone through soldering and counts the seconds for him.

6. Someone says he got promoted, and Griffin shows how it feels about it in its face, voice and gestures at once.

7. Why it is not public yet, with a link to the research preview.

Each demo stands alone, which is what lets each one be quoted on its own, and it is why the quote posts were timed to the demos. Seven posts, seven separate reasons to engage, spread across the afternoon.

The safety hold at the end does the job a waitlist does on other launches: it turns wanting to try it into a reason to follow. [Give them something to reply to](https://doomers.ai/guides/launch-conversation-design) is the mechanism, and [the waitlist guide](https://doomers.ai/guides/waitlist-launch-strategy) is what the days after a gated launch are for.

## What to watch

Day two is where most launches stop, and this one was still adding about a million views every two hours when its second day began. The directory reads the post every half hour until 4 October, then daily, and every figure on this page follows it.

Quotes and replies landed in the first hour here, which is the amplification showing. Bookmarks and reposts land over days, so the saves and reposts rows should be read again at day fourteen before they are treated as settled.

Rank is the other thing. A launch with this much conversation keeps climbing the directory for days as the video keeps being served, and where it stops is the durable result. The figure that matters most to Tavus is the follower count, because it is the audience the public release will launch to.

## The Tavus launch post

[Tavus](https://doomers.ai/launches/tavus-griffin)

@tavus · 43,586 followers

[Open on X →](https://x.com/tavus/status/2105704169009246248)

Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human.

Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. It's the first Human Interaction Model (HIM).

39,121 likes 4,269 reposts 26,913 bookmarks 1,849× its following

[See the full numbers on the launch page →](https://doomers.ai/launches/tavus-griffin)

How Doomers AI launches work

The creators in a Doomers AI launch are paid for their posts, and every compensated post carries X's paid partnership label. [Our disclosure policy](https://doomers.ai/terms#disclosure).

## Questions we get asked

### What is Tavus?

Tavus is a San Francisco company building conversational video AI, what it calls human computing: models that can see, hear and speak with a person in real time, sold to developers through an API. Griffin is its newest model and the first it calls a Human Interaction Model.

### Who ran the Tavus Griffin launch?

Doomers AI amplified it. More than 150 founders, creators and builders from the Doomers AI network were matched to the brief, and the launch post of 1 October 2026 reached 21.3M views.

### What is a Human Interaction Model (HIM)?

Tavus's term for "a new class of model designed to understand and generate face-to-face real-time human interaction": a full-duplex video-to-video model that perceives, decides when to respond and generates speech and video at the same time. Griffin is the first one.

### What is human computing?

Tavus's name for computing that adapts to the person, through face-to-face conversation in voice, video and emotion, instead of making the person learn an interface. Griffin is the model built to be that interface.

### How did Griffin score on NVIDIA's VideoFDB benchmark?

3.83 on generation against a 3.92 human reference and 2.80 for the next-best system, and 3.73 on perception against 4.20 for humans and 3.44 for the strongest baseline. On takeover-rate alignment it scored 62.8% and 73.8%, the highest of any system.

### What is Tavus Griffin?

A full-duplex video-to-video model that Tavus calls the first Human Interaction Model. It listens, watches and speaks at the same time and generates the whole video frame, so what it sees can change what it says while it is saying it. It was announced on 1 October 2026 as a research preview.

### Did Griffin pass the Turing test?

Tavus says it passed a video Turing test: in a blind study of 54 people on one-minute video calls, 48% believed they had spoken to a real human, against a maximum of 2% for earlier systems, including Tavus's own previous stack. The study is Tavus's own and is described on its Griffin research page.

### Can I use Griffin?

Griffin-Lite, a research preview, is open to a select group of early testers. Tavus says a wider release of a more powerful model will follow once further alignment and safety work is done.

### What did Doomers AI do on the Tavus launch?

The amplification. Tavus ran the launch with partners it brought in, and Doomers AI matched more than 150 founders, creators and builders to the brief from every angle, developers, founders, AI model accounts and well-known tech generalists, then timed how each engaged: quote posts with a take in the first hour, and more as each demo landed.

The goal was timeline dominance, the post meeting everyone in tech from more than one direction in the same hours. The timeline answered with more than 30,000 organic posts about Griffin in the first day.

### How is the Griffin launch doing on X?

The post stands at 21.3M views from an account that had about 11,500 followers when it posted, with 5,643 quote posts and 26,913 bookmarks, ranked 36 of the 2512 AI launches we track. It passed a million views inside three hours and ten million inside fourteen. The figures on this page update every half hour while the launch is moving, then daily.

### How many followers did Tavus gain from the Griffin launch?

The company account had about 11,500 followers when it posted and 43,586 a day later, more than 23,000 added in the first nineteen hours, at roughly 1,750 new followers per million views. The timeline produced more than 30,000 organic posts about Griffin in the same period.

### Who backs Tavus?

About $64M in total. Sequoia led the $6M seed in 2023, Scale Venture Partners led the $18M Series A, and CRV led the $40M Series B in November 2025 with Sequoia, Scale and Y Combinator taking part. Tavus came through Y Combinator.

## Keep reading

Tavus Griffin: the launch, taken apart

The post against Tavus's own last launch, the thread post by post, and what a company this size can copy.

Tavus Griffin in the launch directory

The post and the numbers, updated from X every hour.

TypeSafe AI and Jev: from stealth to the largest launch we have run

A model launch from a 1,500-follower account, and what it built.

Give them something to reply to

Why a checkable claim earns quotes, and quotes earn reach.

## Launching a *model?*

We will map the launch window, shortlist the creators who match your buyer profile, and show you the reach math before you commit to anything.
