ModelGo vs OpenRouter for MiniMax H3: Video Outputs, Specs, and Costs

What does MiniMax H3 produce when the brief looks like a real production job? We tested four distinct workflows on ModelGo, recorded $3.011765 in settled charges, and show the outputs, prompts and integration details below. OpenRouter outputs are shown where available, with specification differences called out.

What Is MiniMax H3?

MiniMax H3 combines multimodal context with joint video and audio generation, with an emphasis on instruction following and commercial visual control. These tests examine fabric detail, product geometry, human interaction and controlled installation motion; delivered specifications are reported separately.

Test Method and Actual Charges

We ran four distinct briefs: two text-only and two reference-driven, using matching prompts, reference files and requested settings wherever both platforms were accessible. ModelGo costs were matched to settled billing records and OpenRouter costs come from completed responses’ usage data; the table shows model-output charges rounded to six decimals, excluding separately generated reference art.

Cost scope: these are model-call charges, not the cash outlay required to fund an account. They exclude recharge/payment processing fees, including the 5.5% recharge fee used in the separate pricing comparison; percentage differences here compare call charges only.

The first two ModelGo jobs were submitted on October 30 and recovered on November 4, 2026; the remaining ModelGo jobs and all OpenRouter comparisons were generated on November 4.

Four Production Examples

Tailored fabric fashion film

Production use: Fashion / brand advertising.

Exact prompt

Create a 10-second luxury fashion detail film with an adult model wearing a charcoal tailored coat. Begin close on the shoulder seam and lapel with soft window light, shallow depth of field and visible wool texture. Seconds 0–4: the camera tracks slowly along the lapel as the model turns only slightly. Seconds 4–7: one natural hand fastens a single large dark button without obscuring the coat construction. Seconds 7–10: pull back to a waist-up three-quarter portrait against a warm-gray studio wall. Keep the same face, coat geometry, button count and fabric texture throughout. No text, brand marks, extra fingers, instant outfit changes or exaggerated wind. If sound is generated, use only subtle cloth movement and quiet room ambience. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

Output specifications differ; no equivalent-spec savings claim.

The returned output specifications differ. Compare the displayed dimensions and audio tracks before treating the charge difference as a purchasing decision.

Cycling helmet airflow visualization

Production use: Sports equipment / product advertising.

Exact prompt

Create a 10-second product film for an original matte-white cycling helmet with six visible top ventilation slots and a black retention strap. The helmet sits on a discreet dark support inside a minimal studio. Seconds 0–4: a slow front-to-side camera move reveals shell thickness and realistic surface texture. Seconds 4–7: a few restrained translucent airflow ribbons enter the front slots and emerge from the rear, following plausible continuous paths without passing through solid shell sections. Seconds 7–10: the ribbons fade and the camera holds a clean side profile with space on the left. No cyclist, certification badges, safety promises, numerical performance claims, text or changing slot count. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

Output specifications differ; no equivalent-spec savings claim.

The returned output specifications differ. Compare the displayed dimensions and audio tracks before treating the charge difference as a purchasing decision.

Camping lantern in rain

Production use: Outdoor products / ecommerce video.

This original reference was used for ModelGo and for the paired OpenRouter request where account access was available.

Exact prompt

Use the supplied original camping lantern as the exact product reference. Preserve its olive metal housing, clear cylindrical diffuser, warm internal light and folding top handle. Create a 10-second scene on a sheltered wooden campsite table at dusk. Seconds 0–4: gentle rain falls beyond the table edge while a slow camera push reveals droplets on the lantern housing. Seconds 4–7: a natural adult hand rotates the existing dimmer once, and the warm light increases smoothly without changing color. Seconds 7–10: the hand leaves and the lantern illuminates nearby folded canvas and a metal cup. Keep the enclosure and handle rigid. No new buttons, logos, IP ratings, safety claims or dramatic storm effects. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

Output specifications differ; no equivalent-spec savings claim.

The returned output specifications differ. Compare the displayed dimensions and audio tracks before treating the charge difference as a purchasing decision.

Kinetic museum installation loop

Production use: Culture / exhibition promotion.

This original reference was used for ModelGo and for the paired OpenRouter request where account access was available.

Exact prompt

Animate the supplied original suspended kinetic sculpture for a 10-second square exhibition film. Preserve all seven brushed-aluminum discs, their relative sizes, thin suspension wires and the circular arrangement. In a quiet white gallery, rotate the discs very slowly at slightly different angles while one broad soft light creates moving elliptical shadows on the rear wall. The camera should make a subtle three-percent push-in and remain level. During the final two seconds, ease the sculpture toward its opening arrangement so an editor can make a near-seamless loop. Keep wires attached, disc count constant and reflections physically coherent. No visitors, text, added discs, impossible intersections or claim that the output is already a mathematically seamless loop. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

Output specifications differ; no equivalent-spec savings claim.

The returned output specifications differ. Compare the displayed dimensions and audio tracks before treating the charge difference as a purchasing decision.

How to Calculate the Saving

For outputs with matching delivered specifications, calculate the reduction as (OpenRouter actual charge − ModelGo settled charge) / OpenRouter actual charge × 100. Use the returned dimensions, duration and audio track when deciding whether two outputs are comparable. This article does not project savings from unavailable or mismatched outputs.

Start with the ModelGo API

Install requests and set MODELGO_API_KEY in your environment. The code below uses the endpoint, model ID and parameters from the first real example, with its exact prompt embedded directly. This example submits an asynchronous video job; polling and download steps are described below the code.

import os, requests

prompt = 'Create a 10-second luxury fashion detail film with an adult model wearing a charcoal tailored coat. Begin close on the shoulder seam and lapel with soft window light, shallow depth of field and visible wool texture. Seconds 0–4: the camera tracks slowly along the lapel as the model turns only slightly. Seconds 4–7: one natural hand fastens a single large dark button without obscuring the coat construction. Seconds 7–10: pull back to a waist-up three-quarter portrait against a warm-gray studio wall. Keep the same face, coat geometry, button count and fabric texture throughout. No text, brand marks, extra fingers, instant outfit changes or exaggerated wind. If sound is generated, use only subtle cloth movement and quiet room ambience. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.'
headers = {"Authorization": "Bearer " + os.environ["MODELGO_API_KEY"]}
payload = {'model': 'minimax-h3',
 'prompt': prompt,
 'resolution': '2K',
 'duration': 10,
 'aspect_ratio': '16:9',
 'generate_audio': False}
response = requests.post('https://api.modelgo.ai/v1/content_generation/tasks', headers=headers, json=payload, timeout=420)
response.raise_for_status()
result = response.json()
print(result)  # Save the returned task ID for polling.

After submission, retain id / task_id from the response and poll GET https://api.modelgo.ai/v1/content_generation/tasks/{task_id} until the task reaches a terminal state. Once successful, download the video using content.video_url from the response.

Audio request versus delivery: the recorded requests on both platforms set the audio flag to false. The returned files nevertheless contain an audio stream. “Audio present” describes the detected file structure, not confirmation of audible sound or successful audio control. The code preserves the actual request; these tests do not establish why the audio flag was not reflected in the delivered file.

For first-frame generation on the native content route, send a content array containing the text part and an image_url part with role: "first_frame".

ModelGo and OpenRouter use different model IDs and media request formats. Reuse the production brief, but adapt the endpoint, reference-image structure, task polling and download handling following each platform’s documented request format.

ModelGo API documentation

Use MiniMax H3 on ModelGo

To use MiniMax H3, the model featured in this guide, visit the ModelGo Models page.

Explore MiniMax H3 on ModelGo

Sources and Further Reading

These references cover model capabilities, interfaces and usage records; the briefs were written independently for this test.

• Official model documentation or family showcase: Model capabilities and creative reference; related family documentation is not a specification guarantee for a different variant.
• OpenRouter API reference: Request conventions and platform differences.
• OpenRouter usage accounting: How usage information is exposed; the amounts in this article come from the recorded test calls.
• OpenRouter model catalog API: Check model IDs and catalog properties before a new integration.
• OpenRouter video generation guide: Media request structure and supported workflow patterns.