What to build with H3 Max Live?

Can you actually build with H3 Max Live today?

H3 Max Live is real and currently experimental: fal says the native-continuity checkpoint shown in its live demos is coming to the API next week.

That distinction is important because two different things are being discussed online under roughly the same name. The public H3 Max API already lets developers generate 5-to-15-second videos at up to 768p, with fal reporting less than three seconds to render a five-second clip. H3 Max Live goes further by keeping a video session running and carrying context from one scene into the next. fal says its experimental checkpoint has native infinite continuity, including audio and visuals, with chat prompts appearing in the stream within seconds.

So some H3 Max Live ideas can be prototyped immediately with chained H3 Max clips, while the cleaner version using fal's native long-running checkpoint still depends on an experimental API. We should treat the current demos as a very strong preview of what builders are about to get, rather than assume every production detail has already settled.

What does H3 Max Live actually let us build that H3 Max could not?

H3 Max Live turns H3 Max from a fast clip generator into the basis for a running audiovisual experience that users can keep changing while they watch.

A normal AI video workflow ends after each request. Someone asks for a 15-second scene, the model makes it, and the application receives a finished file. H3 Max Live keeps more of the previous scene in context, which means the next instruction can continue the same world instead of rebuilding it from scratch.

fal's current demo makes the difference easy to understand. Someone watching the broadcast can type a new instruction into chat and see the stream move toward that instruction seconds later. The characters, visual style and previous action can carry forward. fal's engineers say the experimental long-form checkpoint holds the thread across scenes directly inside the model instead of relying on the common trick of taking the last frame and feeding it back into another image-to-video request.

That opens up products where video becomes an output of software state. A vote changes the story. Winning a game changes the next scene. A character remembers what happened earlier. A viewer sends an instruction and the world reacts.

That is a much bigger design space than “generate me a video.”

Is H3 Max Live fast enough for real-time apps?

H3 Max Live is already fast enough for chat-directed shows, turn-based games and slower interactive worlds; frame-by-frame action still belongs to faster world models built specifically for game controls.

fal reports that standard H3 Max can create five seconds of video in under three seconds. Pieter Levels reported roughly nine seconds for a 15-second clip while building Infinite Slop. In both cases, generation can finish before the viewer has watched the footage already sitting in front of them.

That gives developers a useful buffer. Imagine the viewer is currently watching seconds 30 to 45 while the model is already generating seconds 45 to 60. The application can collect the next vote or prompt during the current scene and have another scene ready by the time playback reaches it.

The limits become obvious when we compare H3 Max Live with models built around physical controls. Skywork's Matrix-Game 3.0 reports up to 40 FPS at 720p and accepts action-conditioned inputs for interactive worlds. Alibaba's Happy Oyster already exposes a Wandering mode with WASD and camera controls alongside a Directing mode for changing a scene with text, voice or images. Those products aim at continuous movement. H3 Max Live currently looks much stronger when the user says “the detective opens the basement door” and gives the system a few seconds to dramatize the result.

Experience	H3 Max Live fit	Why
Audience chooses the next scene	Excellent	Seconds of latency barely hurt the experience
Interactive TV show	Excellent	Decisions naturally happen between scenes
AI game show	Excellent	Rules and state can live outside the video model
Choose-your-own-adventure	Strong	Each choice can trigger a new scene
Slow narrative game	Strong	The generation delay can become part of the pacing
WASD open-world game	Weak	Players expect immediate physical control
Competitive action game	Very weak	Frame-level response matters too much

What does Infinite Slop prove about H3 Max Live?

Infinite Slop shows that people will already gather around an interactive AI video stream and try to influence it at meaningful scale.

Pieter Levels launched Infinite Slop within days of H3 Max appearing. He first reported 37,000 viewers, and the stream has since crossed 2,000 concurrent viewers. More interestingly, the product changed almost immediately after people started using it. Levels switched the stream from 16:9 to 9:16 after seeing that most traffic came from phones, then added a queue, upvotes and a “playing next” state. The highest-voted prompt now gets pulled into generation, with the system producing roughly four 15-second videos per minute.

Those changes tell us more than the raw audience number.

Once thousands of viewers shared one stream, the real problem became deciding who controls what happens next. The queue created scarcity. Upvotes created competition. “Playing next” created anticipation. Scrolling backward started turning generated scenes into something closer to a feed.

We would therefore copy the participation loop before copying the endless-video format. The interesting behavior is people fighting over the future of the stream.

Will people actually watch infinite AI video for long?

Pure infinite AI TV looks weak today unless viewers have another reason to keep checking what happens next.

We have already seen endless generative entertainment before H3 Max Live. Nothing, Forever launched as a permanently running AI-generated Seinfeld-style sitcom. TechCrunch reported that the stream had been averaging roughly four concurrent viewers before suddenly going viral through Reddit; during the viral wave, more than 15,000 people were watching at the same time.

That example is useful because the show could already run forever. Endless supply existed from day one, yet almost nobody watched until the project itself became an internet event.

Infinite content therefore does very little on its own. TikTok works because recommendation keeps finding something personally interesting. Twitch works because viewers care about the streamer, competition or community. Television works because characters and events create anticipation.

H3 Max Live needs one of those hooks too.

The best formats we found create a question in the viewer's head: Will our prompt win? Who gets eliminated? What happens if everyone votes for the same ridiculous action? Will this character remember what we did yesterday?

Once that question exists, infinite generation becomes useful because the show never runs out of possible answers.

Why do shared H3 Max Live experiences make more sense than personal ones?

Shared H3 Max Live streams make far more sense than giving every user a private continuous video stream at current video-generation prices.

fal currently lists a 15-second H3 Max 768p generation at $1.20 at its standard rate, with an introductory price of $0.60 currently running. The normal rate works out to $0.08 per generated second.

At the standard price, one uninterrupted hour of generated footage costs about $288. That sounds terrible for one user. The exact same $288 looks very different when 2,000 people watch the same stream: generation cost falls to roughly $0.14 for each concurrent viewer-hour.

Personalization destroys that advantage. A ten-minute personal session contains 600 seconds of generated video, or roughly $48 at the standard H3 Max rate. One thousand simultaneous users doing that separately would create about $48,000 of video generation.

Live pricing may eventually differ from today's normal H3 Max API, and inference prices will probably keep falling. For now, however, the economic direction is unusually clear: one generated world shared by thousands of people is much easier to make work.

H3 Max experience	Cost using current 768p standard rate	What it means
One 15-second scene	~$1.20	Cheap enough for individual experiments
One continuous hour	~$288	Expensive for one viewer
One hour shared by 2,000 viewers	~$288 total	~$0.14 per concurrent viewer-hour
One personalized 10-minute session	~$48	Hard for a free consumer app
1,000 personalized 10-minute sessions	~$48,000	Very hard to subsidize today

Could H3 Max Live power an audience-controlled game show?

An audience-controlled H3 Max Live game show is probably the strongest thing we could build right now.

The format fits the technology almost perfectly. Five generated contestants enter a game. Every 30 seconds, viewers vote on what happens next. “Release a bear.” “Give contestant three immunity.” “Turn the floor into lava.” The winning choice is passed through an LLM director, which converts the vote into a clean scene instruction for H3 Max Live.

Scores, inventories, eliminated players and game rules would live in ordinary software. H3 Max Live would handle the visible consequence. That keeps the game coherent even when generated video gets weird.

We have unusually strong evidence for the collective-control part. Twitch Plays Pokémon let viewers control one Pokémon Red game through chat. Twitch's own final statistics recorded 1,165,140 people entering more than 122 million commands over 16 days, with 121,000 peak viewers and more than one billion minutes watched. The video itself was an old Game Boy game. People came for the chaos of trying to control it together.

Generative AI has also shown that unpredictable outcomes can sustain huge consumer games. Death by AI reached 10 million players in its first month and 20 million within three months, according to its developer's case study with Inworld. Players describe how they intend to survive absurd situations and an AI decides the outcome.

H3 Max Live can make that outcome visible. The audience says what should happen, software decides what the instruction means, and a generated scene shows the consequence.

That feels like a real game rather than a model demo.

Could H3 Max Live make an AI reality show people follow every day?

A persistent H3 Max Live reality show could be even stickier than a one-off game show because recurring characters give viewers something to follow between interventions.

Imagine eight generated people living inside one house. The application keeps their names, personalities, relationships, alliances and memories in structured state. H3 Max Live handles the scenes. Viewers periodically decide who goes on a date, who receives secret information, who gets punished, who enters the house and eventually who leaves.

The key is persistence. A random prompt stream can be funny for five minutes. A character viewers have watched for five days can create attachment, fandom and arguments.

We already have a good AI-native example outside video. Wishroll's Status lets users create and participate in generated social worlds. Inworld reports more than 500,000 daily active users and average daily playtime of one hour and 36 minutes. The product gives people persistent characters and social situations rather than an isolated chatbot conversation.

H3 Max Live could take that basic attraction and turn it into shared television.

We would still keep the canonical memory outside the video model. If Sofia hates Marcus, the application should store that relationship directly. H3 Max Live then receives the relevant memories whenever those characters meet. Native continuity helps scenes flow visually, while software keeps the story true over days or weeks.

Model mistakes may even help this format. Reality TV already thrives on bizarre moments, awkward continuity and unexpected behavior. A strange generated scene can become part of the show's lore instead of ruining a precision-sensitive workflow.

Could people pay to control an H3 Max Live stream?

Paid control of an H3 Max Live stream looks plausible enough to test immediately, especially when control is scarce and everyone can see who won it.

A simple version could keep voting free while auctioning occasional “wildcard” scenes. Viewers submit ideas. One winning prompt comes from votes, then every tenth scene goes to the highest bidder. A brand, creator or community could temporarily take over the story.

There is no direct proof yet that people will pay meaningful money to control generated television. We do have a fresh example showing how much people can spend for scarce, public internet attention. Outbid.lol currently has roughly 1.38 million visitors, and the highest all-time leaderboard position has been bought for $17,000. Its top three positions currently represent $47,000 of bids between them. The product is almost absurdly simple: paying more gives you a more visible position.

Hundreds of copies appeared within days, and a tracker focused on the category now counts more than 300 pay-to-rank boards. Most copies have made very little, which is another useful lesson: scarcity plus an audience can be valuable; copying the mechanic without the audience is worth almost nothing.

An H3 Max Live version gives buyers something more entertaining to compete over. They would be buying a temporary ability to change what thousands of people are watching.

We would test this carefully rather than build the whole business around it, but the experiment is cheap to understand: open a paid slot every few minutes and see what viewers actually bid.

Are brands a better first market for H3 Max Live?

Brands may be the easiest paying customer for H3 Max Live today because a short interactive event can absorb inference costs that would break a free consumer app.

Picture a sneaker launch where viewers direct a generated runner through Tokyo, Mars and a medieval castle. A movie studio could let fans influence a character inside the film's universe. A musician could release an album through a two-hour generated visual world where fans choose each new setting.

The economics become much friendlier in that context. Even several hundred dollars of generation can be small next to the budget for a campaign, event or video production. The brand also brings its own audience, which removes one of the hardest problems faced by a brand-new consumer entertainment product.

H3 Max Live's delay fits these events surprisingly well. Nobody expects a fan vote to change a branded livestream in 50 milliseconds. A ten-second transition between “choose the next world” and seeing that world appear can actually make the vote feel more consequential.

Moderation is easier too. Instead of giving strangers a completely open prompt box, the application can collect suggestions, filter them through an LLM and generate only scenes inside an approved visual universe.

We would expect branded interactive events to appear quickly once H3 Max Live becomes easier to access, even if consumer products take longer to find a sustainable business model.

Can H3 Max Live make personalized interactive movies?

Personalized H3 Max Live movies are technically compelling and economically awkward today.

The product is easy to imagine. A user chooses a character and a world, then watches a movie written around them. Every few scenes, the system asks what they want to do. Each decision sends the story onto a path nobody had to pre-render beforehand.

Generative storytelling already has demand behind it. As seen above, Death by AI reached 20 million players around AI-generated outcomes, while Status has reached more than 500,000 daily users around generated social stories. Those products show that people enjoy stories where their own choices change what happens.

Video makes the cost much harder.

Twenty ten-second scenes would create 200 seconds of footage. At H3 Max's standard 768p price, generation alone comes to roughly $16. A longer 30-minute experience gets into a completely different cost bracket if most of the runtime is freshly generated for one person.

That could still work as a $20 or $30 premium experience, an amusement-park product, a personalized children's story sold occasionally, or a high-end creator product. The economics currently look rough for an ad-supported mobile app where millions of free users expect hours of personalized video.

We would revisit this category aggressively as prices fall. Today, shared stories have the easier path.

Can H3 Max Live replace Unity or Unreal for AI games?

For games, H3 Max Live works best as a cinematic renderer layered on top of ordinary game state rather than as the system responsible for precise movement and physics.

A normal game needs to know exactly where every player, object and collision sits. If someone has 47 health points and three bullets, those numbers need to stay correct. H3 Max Live is built to generate convincing audiovisual scenes from context, so we would let software handle the facts and ask the model to dramatize them.

The market already gives us a useful comparison. Matrix-Game 3.0 was trained specifically around action-conditioned world generation and reports 720p output at up to 40 FPS, including long-horizon memory. Happy Oyster goes even closer to game controls: fal currently offers an exploration mode with WASD and camera movement plus a separate directing mode for text, voice and image instructions.

H3 Max Live's opportunity sits higher up the control stack.

“Move 20 centimeters to the left” is a physics-engine instruction.

“You open the door and discover the king has been replaced by an alien” is an H3 Max Live instruction.

That makes AI RPGs, narrative games, party games and interactive adventures much more natural starting points than shooters or racing games.

Which obvious H3 Max Live ideas should we avoid today?

The weakest H3 Max Live ideas are the ones where expensive generated video adds very little to what the user already gets from text, voice or pre-rendered footage.

A customer-support avatar is an easy example. Customers mostly want the right answer quickly. Continuously generating a photorealistic employee makes that interaction more expensive and introduces extra latency.

Ambient video has a similar problem. A generated fireplace could run forever, yet a normal one-hour fireplace video can loop indefinitely at almost zero marginal cost. Infinite generation solves a problem the viewer barely has.

News is worse. A video model generates plausible images, while news needs footage that accurately represents events that really occurred. A beautifully generated false scene can make the product more dangerous rather than more useful.

Tutoring gets more interesting only when the environment contributes to the lesson. A generated talking teacher offers limited extra value over an avatar. A Spanish learner entering a generated restaurant, ordering food from a waiter and then being sent through airport immigration uses the video world as part of the exercise.

We would use that test repeatedly: if removing the generated world barely changes the product, H3 Max Live probably does not belong there.

Idea	Main issue today	Our view
AI customer-support avatar	Video adds cost without much extra utility	Poor
Infinite generated fireplace	Pre-rendered loops already solve it	Poor
Generated breaking-news footage	Factual visual accuracy is essential	Very poor
Generic talking AI companion	Voice already solves most of the interaction	Weak
Interactive language simulation	Environment changes the learning experience	Promising
Narrative game	Generated consequences add something new	Strong
Audience-controlled broadcast	Many viewers share one generation stream	Very strong

What should we actually build with H3 Max Live?

We would build a shared AI reality game where an audience controls a persistent cast and H3 Max Live renders the consequences.

Start with five or six memorable characters and one clear premise. Put them on an island, inside a mansion, aboard a spaceship or in charge of a failing company. Every 30 to 60 seconds, give viewers a decision. The audience votes. An LLM director turns the winning choice into a scene. H3 Max Live generates what happens.

Keep scores, relationships, objects, money and memories in normal software. Let characters accumulate history. Give viewers accounts so their votes and prompts matter over time. Create recurring moments when one person gets more control than everyone else.

Then add scarcity gradually. Free viewers vote. Active users earn stronger votes. Subscribers can submit special events. Occasionally auction one wildcard scene. Brands can sponsor challenges once there is enough audience.

That format lines up with almost every useful thing we found. H3 Max is already generating faster than playback. fal's experimental Live checkpoint carries context across scenes. Infinite Slop has crossed 2,000 concurrent viewers and is already evolving toward voting and queues. Twitch Plays Pokémon showed how huge collective-control entertainment can become. AI games such as Death by AI and Status show that people can spend serious time inside stories whose outcomes are generated on the fly.

We would also design the product so H3 Max Live can eventually be swapped for another video model. Real-time world models are moving quickly: Matrix-Game, Happy Oyster and fal's broader World Model Accelerator already show several different approaches. The lasting asset would be the characters, audience, game rules, accumulated history and distribution around the show.

H3 Max Live has opened a new category, but “infinite video” is too small a way to think about it. The stronger opportunity is television that behaves like multiplayer software: viewers act, the world responds, and nobody knows exactly what the next scene will be.

What we would build	Potential	Why
Audience-controlled AI reality game	9.5/10	Persistent characters + shared control + stakes
AI game show controlled by viewers	9/10	Simple rules fit the generation delay perfectly
Shared AI RPG or adventure	8.5/10	Generated consequences create huge possibility space
Pay-to-control live channel	8/10	Scarcity can become part of the entertainment
Interactive branded event	8/10	Brands can tolerate higher inference costs
Personalized interactive movie	6.5/10	Great experience, expensive per user today
Educational world simulation	6.5/10	Strong when the environment itself teaches something
Passive infinite AI TV	4.5/10	Endlessness alone gives people little reason to return
Continuous AI companion video	4/10	Voice delivers most of the value much more cheaply
Competitive action game	2.5/10	H3 Max Live is poorly matched to frame-level controls