How to use H3 Max: a source-backed API guide

Start with the linked model records. Match the provider, input mode and pricing variant to your intended job. This example targets Turbo text-to-video; do not assume that changing a name makes it valid for image input or a different service.

Copy the endpoint key from the Turbo model record into a server-only H3_ENDPOINT environment variable. Check the linked official API documentation before running anything. H3 Max Stream does not provide an API key, host a proxy or generate a clip.

Open Calculator, select that endpoint and resolution, then set the number of clips you actually intend to generate. Review the rate's check date, promotion conditions and excluded costs. An estimate is not a spending limit enforced by the provider.

Our recommended first experiment is one authorized, non-sensitive prompt. Decide your maximum spend before running it. Do not begin with a loop, a batch of retries or an unattended stream. A successful clip would still not establish production reliability.

The documented Node workflow uses @fal-ai/client and a FAL_KEY environment variable. Install the client in your own separate server-side example project, not in this website. Never use a NEXT_PUBLIC variable for the key, paste it into a shared URL, or commit it to Git.

The following example is an original, unexecuted adaptation. It intentionally omits a live credential and reads the endpoint from your environment. Inspect current documentation and account billing before executing it.

• In a separate Node.js project, install the documented client with npm install @fal-ai/client. This website does not install or run it for you.
• Set FAL_KEY through your server environment's secret configuration and H3_ENDPOINT to the exact Turbo text-to-video endpoint key from the linked model record. Do not put either value into this site's URL.
• Save the example as example.mjs so Node treats it as an ES module. Review the request and budget, then run node example.mjs only if you intend to send a billable request.
• The sample prints an identifier, not a demo video. Inspect result.data privately in your own program to retrieve the documented output; do not publish signed or private URLs in logs.

import { fal } from "@fal-ai/client";

const endpoint = process.env.H3_ENDPOINT;
if (!endpoint || !process.env.FAL_KEY) {
  throw new Error("Configure server-only H3_ENDPOINT and FAL_KEY first.");
}

// Running this request may incur charges. This guide has not run it.
const result = await fal.subscribe(endpoint, {
  input: {
    prompt: "A paper sailboat drifting across a quiet blue pond.",
    duration: 5,
    resolution: "480P",
    prompt_expansion_mode: "balanced",
    enable_safety_checker: true,
  },
});
// Inspect the returned video privately; do not log credentials or prompts.
console.log(result.requestId);

The documented response contains a video file; inspect it in your own environment. Check whether it answers the prompt, whether the content is appropriate, and whether actual charges match your assumptions. Keep provider documentation claims separate from your own observations.

For a batch workflow, define a manual acceptance checklist before increasing volume: usable output, permitted content, acceptable total delay, and affordable cost including failed attempts. Record uncertainty rather than declaring a speed or quality benchmark from a single example.

These are editorial debugging recommendations, not a catalog of guaranteed provider error codes. First determine whether a request was accepted. If you retained an identifier, consult the documented queue status/result workflow before creating another request.

• Input rejection: compare field names, case-sensitive values and input mode against the selected endpoint's current schema.
• Authentication trouble: inspect your server environment privately. Do not send the key in a screenshot, support message or browser query.
• Timeout or unclear outcome: avoid a retry loop. Check the existing request and account history; contact provider support when its outcome remains uncertain.
• Safety rejection or unsuitable output: revise the task to permitted content. Do not disable safety checks to force a result.
• Unexpected cost: stop further submissions and compare the actual endpoint, resolution, current rate and request count with your estimate.