What Infinite Slop Actually Costs to Run on H3 Max

Pieter Levels shipped a channel that generates its own programming

Pieter Levels, the solo builder behind Nomad List and Remote OK, launched Infinite Slop in the last week of August 2026. It is a livestream with no fixed footage: every clip is generated on the spot by an AI video model, chained one after another, with the chat picking what comes next by typing a prompt. Nobody edits it, nobody schedules it, and nothing about it repeats unless the chat asks for the same thing twice.

The reason it works as a live channel instead of a slideshow of loading spinners is speed. The model behind it, MiniMax H3 Max, is fast enough that a five-second clip renders in under three seconds on fal's infrastructure, according to Steal What Works' reporting on why Levels built it now rather than a year ago. That gap between render time and playback time is the entire trick: the channel can generate the next clip while the current one is still playing, so viewers never see a loading screen.

Fal's H3 Max is a post-trained, speed-tuned build of MiniMax's own model

H3 Max is not MiniMax's stock model. Fal describes it on its own model page as "a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality." In practice that means fal retrained and re-served the model specifically to be fast enough for exactly this use case: continuous generation with no visible seams.

Latent Space's AINews newsletter covered the launch the same week, describing Infinite Slop as an example of "the infinite videogen loop" becoming viable now that render time has dropped under real-time playback speed for short clips. That is a new category of product, not a faster version of an old one: a livestream where the content only exists because someone is watching it.

If you are tracking which model releases actually change what is buildable rather than reading launch threads for the sake of it, Clawsmith tracks demand signals like this one as they happen instead of after the fact.

Why this format did not exist a year ago

Continuous AI video is not a new idea; live, chat-directed AI video that never stalls is. Most video models on fal and elsewhere still quote generation time in the range of 30 seconds to several minutes for a five-to-ten-second clip, which is fine for a batch job and unusable for a live feed, because a viewer watching a spinner for a minute between five-second clips is not watching a livestream. H3 Max's under-three-second render for a five-second clip is roughly the first time on fal's own catalog that generation time has dropped meaningfully below playback time for a usable resolution, which is the actual unlock, not the aesthetics tuning fal also mentions on the model page.

What running a channel like this actually costs, priced off fal's own page

Fal prices H3 Max text-to-video by the second of output, not by the clip. As of today, the model's own pricing panel reads: $0.0125 per second at 480p, $0.02 per second at 768p. Fal labels these promotional launch rates, 75 percent off the standard price, ending September 7, 2026. After that date the same page states 480p goes to $0.05 per second and 768p to $0.08 per second, a flat 4x increase on both tiers.

Because the price is linear per second, the cost of running one continuous channel does not depend on how many clips it is cut into, only on total seconds of video generated. A single channel running 768p video nonstop for 24 hours needs 86,400 seconds of generated footage:

• At the current promotional rate: 86,400 seconds x $0.02 = $1,728 per day.
• At the post-September-7 rate: 86,400 seconds x $0.08 = $6,912 per day, the same 4x jump fal states on its own pricing panel.
• Dropping to 480p cuts both numbers by 37.5 percent: $1,080 per day now, $4,320 per day after the discount ends.

That is the cost of ONE always-on channel, not per viewer. A chat-directed livestream like this one does not re-render for each person watching; everyone sees the same generated feed, so the bill does not scale with audience size the way a per-user AI feature normally does. That is the structural reason this format is cheap enough for a solo builder to run at all: the cost curve looks like a media company's bandwidth bill, not like an AI SaaS product's inference bill.

The build is small enough to describe in five steps

Nothing about the mechanism is exotic. The loop fal's own docs and the coverage above describe is:

1. Collect the next prompt. Infinite Slop reads it from chat; a solo build could pull from a queue table, a Discord channel, or a fixed rotation of prompts.
2. Call the H3 Max text-to-video endpoint (minimax/h3-max/text-to-video on fal) with the prompt, a duration, a resolution (480p or 768p), and an aspect ratio.
3. Poll for the result. Fal's API returns a job you poll until the video is ready; at H3 Max's speed that wait is close to the clip's own length.
4. Queue the finished clip behind whatever is currently playing, so there is always one clip in the player and one clip rendering behind it.
5. Loop back to step 1 before the currently playing clip ends, so the queue never runs dry.

The only genuinely hard part is step 4: keeping exactly one clip ahead at all times so playback never stalls waiting on a render. Everything else is a standard API call against a documented endpoint, priced per second on a public page.

What disappears when nobody is recording it

Infinite Slop itself does not keep a public archive. Once a clip finishes playing, it is gone from the feed; the only place to see the model's own live output continuously, with every clip and the prompt that produced it kept on record, is a restream at theinfiniteslop.com, which mirrors the same feed and logs each clip as it airs rather than discarding it after one playthrough.

The number that actually matters is the date, not the price

$0.02 a second sounds abstract until you multiply it by a day of continuous output, and $1,728 a day sounds like a lot until you compare it to $6,912 a day three days later. Fal's own pricing page puts the discount's end at September 7, 2026, which means anyone testing whether a continuous AI-video format is viable for their own idea has a real, dated reason to run that test this week rather than next month. The math does not change gradually; it steps 4x on a fixed date fal has already published.

Summary

Infinite Slop runs on fal's H3 Max, a speed-tuned build of MiniMax H3 that renders a five-second clip in under three seconds. Fal prices the model per second of output: $0.02 at 768p and $0.0125 at 480p today, rising 4x to $0.08 and $0.05 on September 7, 2026. A single 24-hour channel at 768p costs $1,728 a day now and $6,912 a day after that date, and that cost does not scale with how many people are watching, only with how many seconds of video the channel generates. The build itself is a five-step loop against one documented API endpoint. The live output, archived clip by clip with its prompts kept, is running now at theinfiniteslop.com.