A weekend after fal launched H3 Max, builders turned five-second AI clips into live TV channels, a 24/7 news station, and games played by prompting.
fal's H3 Max generates AI video faster than real time, turning generation from a slow render into a live medium builders can prompt like a chat room.
MiniMax open-sourced its H3 video model, and fal used it to build H3 Max, which generates a 5-second 768p clip with audio in under 3 seconds, about 35x faster than MiniMax's own official endpoint. That speed turned video generation from a slow render into something people could prompt live: fal Live streams six chat-directed channels, Levels.io built a thousand-viewer 'Infinite Slop' channel from his phone in a sauna, and other builders shipped a 24/7 AI news station, an interactive Seinfeld-style mystery, and a chat-voted choose-your-own-adventure game. It's free to try, costs about $4.80 a minute at the full API rate, and nobody has confirmed whether the model itself will be open-sourced.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →
Cold open: real-time AI video generation has arrived, opening up new use cases the video will walk through, including whether it will be open-sourced and what it costs.
Two predictions set up the moment: LTX's CEO Zeb Farbman said video generation faster than you can prompt was coming soon, and Nvidia's Jensen Huang said every pixel will eventually be generated.
The wave exists because MiniMax open-sourced its H3 video model, which has become the model the open-source video community has rallied around.
fal built H3 Max on top of open-source H3 for near real-time generation; a demo teaser clip (an interactive Seinfeld-style bit) previews what's coming later in the video.
fal is introduced as the inference platform that powers many generative-AI apps behind the scenes, effectively the shared infrastructure other products are built on.
H3 Max renders a 5-second 768p clip with audio in under 3 seconds, about 35x the throughput of MiniMax's own H3 endpoint and roughly 15x faster than comparable-quality models.
The fast/cheap/good tradeoff in practice: a locally-run Turbo LoRA version of H3 (via Matt VidPro) edged out H3 Max slightly on quality in a side-by-side 'crab with a jetpack' test.
H3 Max is free to try on fal with no signup: a number of free 5- or 10-second text-to-video or image-to-video generations per day, usually finishing in under 30 seconds including queue time.
fal Live streams six channels (Chaos Stream, Toon Lab, Choice, Tiny World, Sitcom, Hot Mic) that generate continuously based on live chat prompts, currently on standby while fal tunes it.
Community builds using H3 Max: James Russo's infinite TikTok-style vertical feed and a Rick and Morty-style interdimensional-cable channel, where randomness reads as intentional rather than broken.
Levels.io built Infinite Slop, a live channel with over 1,000 concurrent viewers, entirely on his phone in a sauna. It runs on H3 Max Director, an autoregressive variant with up to two minutes of native context so characters and settings persist across separate generations.
Igor Sack built Channel Zero, an always-on AI news channel fed by RSS feeds, Reddit, and a rooftop antenna, turning real-world data straight into generated video.
Chat-directed prompts drive an actual branching mystery: viewers vote on interrogation choices in a Seinfeld-style detective storyline about a stolen soup recipe book, built entirely from 5-second generation limits.
Using real, recognizable characters like the Seinfeld cast raises the obvious copyright question the video flags directly as something the industry still has to work out.
More structured interactive builds: Ege Bashe's real-time AI Pictionary game, and Blendy's branching choose-your-own-adventure game, with the code open-sourced on GitHub for others to fork.
API pricing is about $0.08/second (roughly $4.80/minute) at full rate, currently half off at $0.04/second for a limited window; running a 24/7 channel like Infinite Slop would cost around $4,000 a day.
No official word yet on whether H3 Max will be open-sourced, though the autoregressive Director layer will likely stay proprietary to fal even if the base model is released.
Closing take: quality isn't at cinematic-AI levels yet and Levels.io's own 'slop cannon' label is fair for now, but the pace of improvement plus the interactivity layer (like a chat-voted D&D campaign) points to a genuinely new entertainment medium, not just faster rendering.
MiniMax's open-sourced H3 model let fal build a generator fast and cheap enough to turn AI video into a live, interactive medium instead of a slow rendering process.
“Video generation faster than you can prompt will be possible very soon.”
“In the future, every pixel will be generated.”
“We always have to kind of go back to the fast, cheap, good triangle. You can only pick two.”
“Is it as Levels.io appropriately named it a bit of an AI slop cannon? Yes, it currently is, but that won't last long.”
“How about a D&D campaign where chat votes on decisions at each point?”
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
Real-time AI video stopped being a promise and started being a product this week: fal's H3 Max can generate a five-second clip faster than you can watch it, and the internet spent exactly one weekend turning that into live TV channels, a 24/7 news station, and games you play by typing.
You can only pick two: H3 Max is fast and cheap, which is why a slower, locally-run alternative can still look slightly better.
00:00
00:1700:25
00:38
00:5001:01
01:14
01:2501:30
01:45
01:56
02:1002:17
02:32
02:4102:50
03:01
03:0903:26
03:37
03:4703:59
04:10
04:2504:37
04:43
04:58
05:0605:16
05:28
05:39
05:50
06:01
06:13
06:24
06:3506:46
06:52
07:08
07:23
07:31
07:42
07:53
08:0408:15
08:30
08:37
08:4909:00
09:09
09:22
09:3309:44
09:55
10:06
10:1810:29
10:45
10:50
11:00
11:13
11:24
11:36
11:4211:58
12:09
12:20
12:26
12:4212:53
13:05
13:1313:27
13:37
13:49
14:00
14:14
14:22
14:3814:42Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
14:21Joseph Martin runs five head-to-head tests between OpenAI's new GPT-6 Astra and the cheaper GPT-5.6 models inside Higgsfield's AI video and image tools, to see if the price jump is worth it.
September 7th
22:37An ex-Rolls-Royce ML engineer wires Claude Code straight into BytePlus's Seedream and Seedance APIs, skips the credit-metered ad platforms entirely, and live-builds a JBL speaker ad for $6.64.
August 29th
13:44One prompt file sent an autonomous coding agent off to research, script, art-direct, and render a full Vox-style explainer video by itself.
September 3rd
06:50A step-by-step walkthrough of building a consistent AI character and a full cinematic clip inside a single ChatGPT Work conversation, using the Higgsfield plugin and Seedance 2.5.
August 15th
25:22A step-by-step walkthrough of reverse-engineering a bestselling Etsy listing, then building a done-for-you Instagram Reel template pack with Claude connected to Higgsfield's AI video MCP.
September 2nd
20:34A workflow for turning Pinterest references into Claude-written Seedance 2.5 prompts, then finishing the render with five CapCut effects that make AI video read as hand-animated.
August 27th