Genie 4, Tavus and Reactor: Generated Worlds, Generated Faces, and What Still Needs a Camera

Yesterday Google DeepMind announced Genie 4, a model that turns a sentence into a world you can walk around in, now with sound. Earlier this year Reactor came out of stealth with $59 million to host models like it. Tavus already sells AI humans that hold a video call with you and answer within a fraction of a second.

Put those three next to each other and you get the shape of where interactive video is going: worlds made on demand, faces made on demand, and the plumbing to stream both to a phone. I build a filmed story that calls you, so I have a stake in this. Below is what each one actually does today, sourced and dated, and then my honest read on what it changes for anyone telling stories people can talk back to.

The short answer

• Google Genie 4 generates explorable worlds in real time from a text or image prompt, and is the first Genie to generate audio, including dialogue. It is a research preview: trusted testers only, with wider access "coming soon."
• Tavus sells real-time AI humans you can video-call. You pay by the minute, and the company says it is used for sales training, tutoring, interviewing and companionship.
• Reactor is the infrastructure layer: an SDK and API for running, streaming and steering real-time video and world models, billed by usage.
• What it changes: improvised, open-ended scenes are now cheap enough to prototype. What it doesn't: a performance with intent, consent and an ending someone wrote still comes from a person in front of a camera.

Google Genie 4: a world from a sentence

DeepMind's Genie line is what the field calls a world model: instead of rendering a scene from 3D assets, it predicts the next frame of video from what you did a moment ago, so the world is generated as you move through it.

Genie 3, announced in August 2025, ran at 720p and 24 frames a second, stayed consistent for a few minutes, and remembered what it had shown you for about a minute. In January 2026 Google opened a consumer version, Project Genie, to US subscribers of its top AI Ultra plan; reporting at the time described sessions capped at 60 seconds, with noticeable input lag.

Genie 4, announced on 27 October 2026, adds three things that matter for storytelling:

1. Native audio. Google calls it its first real-time world model to generate sound itself, including music, ambience and interactive dialogue.
2. Steering mid-scene. TechCrunch reports you can change things like weather or camera while the simulation runs, and that a scene holds together for up to two minutes, double Genie 3.
3. Labelling. Every frame and sound carries Google's SynthID watermark.

Google is candid about the limits. Its own post promises coherent environments "over short sessions" and lists complex physics, accurate real-time audio, legible text and long-term consistency as unsolved. It has published no resolution, frame rate or price for Genie 4.

What it means for stories: a world model is a stage, not a script. It is superb at "what is around the corner" and has no idea why you should care about the corner. Two coherent minutes is a scene, not a story, and nothing in the model knows how a scene should end.

Tavus: a face that answers

Tavus builds what it now calls PALs, AI humans that see, hear, remember and respond over video, voice and text. The company raised a $40 million Series B in November 2025 led by CRV, and launched PALs alongside it.

Under the hood are separate models for each part of a conversation: one renders the face (Phoenix), one reads your expression and surroundings (Raven), and one manages turn-taking so it neither talks over you nor leaves dead air (Sparrow). Its homepage currently claims 134 millisecond rendering latency and more than 150,000 developers, with customers using it for sales role-play, clinician education, job interviews and tutoring.

It is billed by the minute. Developer plans start free with a small monthly allowance, and the pricing page lists overage in the range of roughly 26 to 37 cents a minute depending on plan.

What it means for stories: this is the closest anyone has come to a character who can genuinely improvise with you on camera. It is also an ongoing cost for every minute of every viewer, which is the opposite of a film's economics: a film costs the same whether one person watches it or a million do.

Reactor: the pipes underneath

Reactor is not a model. It is the layer that runs models like these and streams them to an app. It launched on 28 May 2026 with $59 million led by Lightspeed, and its founders previously led technical work on Apple Vision Pro.

On its homepage Reactor lists a catalogue of hosted real-time models, from explorable worlds to a portrait that speaks a script live, and claims frames arrive in under 50 milliseconds. Developers reach it through an SDK and pay by usage.

What it means for stories: the hard part of real-time generated video used to be serving it fast enough to feel live. A platform whose whole job is that problem means small teams can try generated scenes without building a GPU cluster first. That is genuinely new in 2026.

Generated versus filmed: where each one wins

Here is my read, as someone who chose the filmed side on purpose.

• Open-ended improvisation. Generated: strong. Filmed: limited to what was shot.
• How long a scene holds together. Generated: minutes. Filmed: as long as the film.
• Cost of one more viewer. Generated: rises with every minute watched. Filmed: close to zero.
• A deliberate ending. Generated: has to be bolted on. Filmed: written, shot and cut.
• Performer consent. Generated: negotiated per likeness. Filmed: it is the performance itself.
• The feeling that someone meant it. Generated: rare so far. Filmed: the whole point.

Generated video wins when the value is range: a world that goes wherever you walk, a tutor who can answer any question. Filmed video wins when the value is intent: a person who chose how to deliver a line, and a story built to arrive somewhere.

Consent is not a footnote here. The 2025 SAG-AFTRA interactive media agreement, ratified after an 11-month strike, requires written, specific consent before anyone makes or uses a digital replica of a performer. Generated characters built from no one sidestep that. Generated versions of a real actor do not.

Where Treezy Play sits

Treezy Play is the filmed side of this map. Dr. Tiffany Belle is played by a real actress, on camera, and partway through the film she video-calls you. You answer out loud, the story listens, and it cuts to the filmed scene that fits what you said: four branches, twenty-eight endings, about forty minutes.

Understanding what you say is where AI earns its place for us. I wrote up how that works, including what happens when it can't decide, in Never Stall. What we deliberately don't do is generate her face or her voice. When the story needs to ask you something again, it asks in a caption, because the performer's consent covers her performance and nothing synthesised from it. That constraint turned out to be a feature: the call never sounds like anyone but her.

What to watch next

1. Genie 4 access. Google says Project Genie for AI Ultra subscribers and early API access are next. Watch for session length: minutes, not seconds, is when it becomes a storytelling tool.
2. Per-minute price. Real-time AI humans become a mass-market story format when a forty-minute session costs cents, not dollars.
3. Hybrids. The interesting work over the next year will mix the two: filmed performances for the moments that need to land, generated material for the open spaces in between, clearly labelled.

FAQ

What is Google Genie 4? A real-time world model from Google DeepMind, announced 27 October 2026, that generates explorable environments with sound from a text or image prompt. It is in research preview with trusted testers.

Can I use Genie 4 today? Not unless you are a trusted tester. Google says access is coming to Google AI Ultra subscribers through Project Genie and to early partners through the Gemini API, without a date or price.

What is the difference between Tavus and Reactor? Tavus makes AI humans you talk to on a video call. Reactor is a developer platform that hosts and streams many real-time video and world models, so apps don't have to run them on their own servers.

Is Treezy Play AI-generated? No. The film and Dr. Tiffany Belle's performance are shot with a real actress. AI is used to understand what you say on the call and choose the filmed scene that follows.