The same post-trained MiniMax H3 Max, served through fal-compatible endpoints. A 5-second 768p clip lands in ~4.5s end-to-end — you just pay half the list price for it. We run exactly one model family — no catalog sprawl, all infrastructure tuned for H3 Max.
Private beta — drop your email, get an API key the moment we launch.
$ protix run --model h3-max-turbo --seconds 5 → billed $0.04 at launch · fal bills $0.20
One model family. MiniMax H3 Max — and nothing else on the meter.
The same MiniMax H3 Max model family — at half of fal's list price. Every price below is per second of output, billed only on success.
| Endpoint | fal · list | Protix · −50% | Protix · launch−80% |
|---|---|---|---|
| H3 Max Turbo480p · t2v / i2v | $0.025/s | $0.0125 | $0.005 |
| H3 Max Turbo768p · t2v / i2vPOPULAR | $0.040/s | $0.020 | $0.008 |
| H3 Max Turbo1080p · t2v / i2v | $0.080/s | $0.040 | $0.016 |
| H3 Max480p · t2v / i2v | $0.050/s | $0.025 | $0.010 |
| H3 Max768p · t2v / i2vPOPULAR | $0.080/s | $0.040 | $0.016 |
| H3 Max1080p · t2v / i2v | $0.160/s | $0.080 | $0.032 |
| Reference-to-Videooutput + ref tokens > 4,096 | $0.080/s + $0.02/1k tok | $0.040/s + $0.01/1k | $0.016/s + $0.004/1k |
| Directorstreaming · 60s minimum | $0.080/s · min $4.80 | $0.040/s · min $2.40 | $0.016/s · min $0.96 |
Reference-to-Video bills output seconds plus reference tokens over the 4,096 free allowance — halved on both sides.
Speed is a property of the model, not the storefront. Same H3 Max — a 5s 768p clip still lands in seconds, not minutes. What changes is the invoice.
In a public 300-run benchmark of H3 Max Turbo (5s · 480p): p50 5.03s, p95 15.42s. The API's own inference median was 0.45s — the tail is queue, encode and download. We quote end-to-end, always.
Prompt expansion ships set to fast on preview endpoints — balanced adds ~1s, quality can add ~30s. You pick the mode per call; the default favors speed.
Queue waits, upstream errors and expired jobs cost $0.00. You pay for seconds of video that actually come back — metered per output second, reported per request.
Endpoints mirror the fal request and response shape. Swap the base URL and key; keep your models, payloads and polling logic.
import { fal } from "@fal-ai/client"; fal.config({ credentials: process.env.FAL_KEY,}); await fal.subscribe("minimax/h3-max-turbo", { input: { prompt, duration: 5, resolution: "768p" },});
import { fal } from "@fal-ai/client"; fal.config({ credentials: process.env.PROTIX_KEY, host: "api.protix.win",}); await fal.subscribe("minimax/h3-max-turbo", { input: { prompt, duration: 5, resolution: "768p" },});
This is the entire catalog. One model family, tuned end to end — that's how it stays this fast and this cheap.