H3 Max, Turbo and Director: choose your model.

fal's launch material describes work on both the model and the system serving it. The model is post-trained for qualities such as prompt adherence and aesthetics, while the inference stack is optimized for speed. These are provider claims; they should not be read as independent results from this publication.

The published five-second example renders in under three seconds. That is the useful threshold for a simple broadcast design: in favorable conditions, a generator can prepare content faster than an audience consumes it.

H3 Max should not be treated as an interchangeable name for every MiniMax H3 endpoint. The provider publishes separate endpoints and capabilities. When building or comparing costs, record the complete endpoint ID and resolution instead of writing only “MiniMax pricing.”

The documented H3 Max interfaces include text-to-video and image-to-video. The latter is useful when a starting visual matters, while the former is a straightforward way to test a scene concept.

The API documentation is the authority for current parameter names, allowed durations and image requirements. Do not assume that a capability on another H3 model is available on H3 Max.

An image input does not establish rights to the image. Use your own, properly licensed or otherwise permitted material. Keep keys on your server; a browser should never receive your private provider credential.

FastH3 and H3 Max are related to MiniMax H3, but they are not two interchangeable endpoint names and the public evidence is not a clean head-to-head benchmark. H3 Max is fal's post-trained variant served through its hosted API. FastH3 Preview v1 is a FastVideo-distilled checkpoint that Reactor wraps in a generation queue and WebRTC playout service. A third route, illustrated by SlopTV, runs MiniMax H3 locally with a Turbo LoRA and uses a replay pool to cover slower generation.

Those timing rows answer different questions. They use different clip lengths, implementations and hardware, so they do not prove that one model has higher output quality or lower total cost. H3 Max's provider benchmark shows substantial generation headroom for short clips. Reactor's four-B200 FastH3 profile is approximately real time for a 14.375-second clip; its eight-B200 deployment is intended to build ahead of playback.

SlopTV demonstrates a slower local alternative rather than another direct benchmark. Its author reports about 90 seconds for one 15-second, 352×608 clip per RTX 5090. Two workers alternate, yielding a new clip roughly every 45 seconds, while fresh-first replays keep the RTMP stream moving. That is a valid continuous channel design even though generation itself is not real time.

Choose from your operational constraint. A hosted H3 Max route avoids running the model but exposes you to provider pricing and limits. FastH3 offers inspectable weights and a specialized low-step path, but the cited real-time profile requires data-center GPUs. A local H3 route can be easier to own end to end while accepting lower resolution, slower freshness or more replays. For a fair test, run the same prompt set and clip specification, then record playable p50/p95 latency, accepted-output rate, cost per accepted minute and recovery after an interrupted job.

H3 Max Turbo adds a separate, lower-priced clip API. Standard 480p/768p rates are $0.025/$0.04 per second; its September 7 launch offer is $0.00625/$0.01. It is distinct from the Turbo LoRA used by some local H3 projects.

At 768p standard pricing, thirty days of uninterrupted fresh output costs $103,680 with Turbo, compared with $207,360 with Max. The Turbo launch scenario is $25,920. These are calculated generation budgets, not the official Infinite Slop bill or a statement that its operator switched models.

For a broadcast, use saved time to build a playback reserve. For a branching story, budget discarded alternatives as well as the selected scene. Test the same prompt set and measure accepted output before deciding that the cheaper endpoint is the better fit.

Director serves a different task: steer a continuous session instead of requesting a sequence of independent files. Its session minimum needs a separate calculation.

The H3 Max text-to-video endpoint lists standard rates of $0.05 per output second at 480p and $0.08 at 768p. The current launch snapshot is $0.0125 and $0.02, ending September 7. This updates the earlier $0.025/$0.04 launch announcement; use the endpoint and your account terms for current billing.

These are arithmetic examples from the September 5, 2026 public-price review. Standard rates remain the calculator default. Use the custom-price field for a different account rate, or the separate Director session calculator for session-based usage.

A five-second scene generated in less than five seconds can create headroom for playback. It does not guarantee an instant response to a new audience suggestion: queueing, moderation, download and your existing buffer also contribute delay.

Measure when a direction is accepted and when the viewer actually sees it. Keep those timestamps separate from model inference. In a branching show, pre-generating alternatives can reduce the wait after a choice, but every discarded branch can add cost. A larger buffer improves continuity while also moving the visible response farther behind live input.

Measure the entire path that your viewer depends on: prompt acceptance, moderation, provider queue, generation, result download and playable delivery. A fast inference measurement can coexist with a slow or unreliable overall experience.

Run several short trials and record the median and slower-tail completion times. If a five-second clip occasionally takes twelve seconds to become playable, a one-clip reserve will not protect the broadcast every time. You may need more buffer, fewer fresh minutes, a scheduled show or reviewed replay material.

Also measure time spent rejecting output. An output that finishes quickly but fails your safety or quality check contributes cost without adding broadcast time. Keep that separate from requests that fail before billing. The calculator's extra-generation percentage represents additional billable work, not a blanket claim that every API error costs money.

Because audio is part of the model's output, short sound instructions can be useful: a spoon striking a glass, gentle room ambience or one brief line of dialogue. Keep the sound consistent with the visible action.

For a short clip, a single conversational line is easier to fit than a long script. Avoid asking for several people to speak at once, and do not assume generated text, pronunciation or synchronization will be perfect. Review the actual result before adding it to a public broadcast.

The prompt studio assembles a scene, action, camera and sound into editable text. It does not call fal, charge an account or imply that its examples have already aired. You decide where to use the prompt.

A model endpoint is one part of the system. You also need a persistent queue, rate limits, prompt and output checks, a buffer that stores only approved clips, and a delivery layer capable of serving your audience.

Put a hard generation budget behind the submit path. A public chat box should not translate into unlimited paid requests. Reserve expected cost before submitting a job, reconcile its outcome afterward and stop accepting work when the ceiling is reached.

For a first experiment, pick a scheduled window and one fictional setting. Use the cost calculator to model it, then follow the build checklist to test interruptions and restart behavior. Increasing hours or resolution is easy; recovering from an unbounded queue is harder.