FastH3 can autoplay. Its default queue does not.

The little setting behind the big claim

Project says: Reactor’s FastH3 implementation has distinct generation and playout queues. After a clip finishes—or after stop—output flushes to black and waits for the next play by default.

The same project documents set_autoplay. When enabled, it starts the ready clip at the front of the playout queue whenever nothing is playing. That is useful, but it changes one narrow thing: who starts an already-finished clip. It does not create a new clip, refill a drained queue, or repair a late render.

The reference client shows the manual path

Project says: Reactor’s reference queue client enqueues two prompts, waits for them to become ready, plays the first, confirms that the stream holds black, then plays the second. The client records the received clips, messages, and timing report.

That is a useful smoke test. It is not a production stream test, an RTMP uptime record, or evidence that viewers accept the black fallback. It does show exactly why a “never-ending” claim needs its queue policy beside it.

Autoplay does not make a continuing story

Project says: FastH3 clips are independently generated and hard-cut; the project says subject, framing, and voice do not carry across the boundary. Autoplay can move from one ready scene to another without waiting for an operator. The next scene is still a new scene.

Use this route for a controlled run of short scenes only if you can state the ready buffer, default and enabled behavior, fallback picture and audio, prompt-to-play delay, and what happens after a failed build.

The more expensive continuity route has a deadline

Provider says: fal’s H3 Max Director holds a WebRTC session and carries context into later segments. Its page lists $0.02 per generated second right now and $0.08 from 14 September 2026, with a 60-second floor and public sessions up to 15 minutes. The stated full-session generated-video spend moves from $18 to $72 at the stated price change.

This is a provider description and price, not a pool measurement. It does not tell us the real direction delay, reconnection behavior, operating bill, or whether a viewer sees a clean handoff.

What to show in the next demo

1. Start with autoplay off and let a clip finish.
2. Turn autoplay on with two ready clips, then show the transition.
3. Drain the queue and show the picture and audio while the next render is late.
4. Show whether the next clip remembers the world or begins a new scene.

That record would tell a creator more than a stream that never shows its failure mode.