H3 Max vs H3 Max Turbo
Both fal variants ran the same fixed prompts at 768P with prompt expansion disabled. The requests were made in separate batches on September 18, 2026. Queue polling is included in the recorded wait; inference time alone is not what we measured.
Anvisha Pai, Co-founder & CEO, Voyager
Which to use
Start with Turbo for low-cost drafts. All nine requests returned in both batches, with estimated fixed-set totals of $0.90 for Turbo and $1.80 for Max at the September launch prices. Sampled frames show readable title text and a visible kettle and filter in both runs. Neither text-only product shot is exact: Max adds a handle and Turbo adds a spout-like lip as well. Max’s single run does not establish a universal quality advantage. Compare the actual shot, use image conditioning for a specific product, and budget at the regular rate if delivery falls after September 30.
Side by side, measured
| H3 Max | H3 Max Turbo | |
|---|---|---|
| Vendor | fal / MiniMax | fal / MiniMax |
| Released | Date not verified | Date not verified |
| Price | $0.04 per second at 768p via fal | $0.02 per second at 768p via fal |
| Seconds to make one clip | 6 to 26 | 6 to 6 |
| 9 clips cost | $1.80 | $0.90 |
| Length | 5 to 15 seconds | 5 to 15 seconds |
Seconds and cost are from our own runs of the same 9 prompts through each model.
The same 9 prompts, H3 Max against H3 Max Turbo
One try each, nothing re-rolled or retouched, generated by us. H3 Max is on the left, H3 Max Turbo on the right. Every prompt has its own page with the whole roster.
Product turntable
Every model on this prompt →Both invent product features. Turbo’s dripper has a spout-like lip; Max’s has a handle. Neither is a faithful catalogue asset from text alone.
›What we asked for
A matte black ceramic pour-over coffee dripper rotating slowly on a white turntable, soft studio light, seamless white background, product video, no text.
Person speaking
Every model on this prompt →›What we asked for
A woman in her 40s with short grey hair stands in a bright kitchen, looks at the camera and says 'Good morning, the coffee is ready', natural window light, handheld feel, she smiles at the end.
Animated title card
Every model on this prompt →Both resolve into readable title lettering after the entrance. The small date line should still be inspected at output size.
›What we asked for
A title card reading 'MODA SUMMER SESSIONS' animates in over slow-motion footage of a festival crowd at dusk, then a smaller line 'Friday 12 June' fades in beneath it. Bold condensed type, tangerine on navy.
Camera move
Every model on this prompt →›What we asked for
Drone shot pushing slowly forward over a coastal town at sunrise, terracotta roofs, fishing boats in the harbour, long shadows, cinematic and steady.
Physics
Every model on this prompt →›What we asked for
A glass of water tips over on a wooden table and spills in slow motion, water and three ice cubes sliding across the wood, macro lens, morning light.
Hands at work
Every model on this prompt →Both keep the kettle and filter in the sampled frames, unlike the earlier base-H3 result.
›What we asked for
Close-up of hands pouring hot water from a gooseneck kettle in slow circles into a pour-over coffee dripper, steam rising, shallow depth of field, warm cafe light.
Busy scene
Every model on this prompt →›What we asked for
A crowded Saturday farmers market seen from slightly above, shoppers with tote bags, a dog on a leash, a busker with a guitar, bunting swaying, late morning light, gentle camera drift.
Specific art style
Every model on this prompt →›What we asked for
A lighthouse on a rocky cliff in a storm, in the style of a Japanese ukiyo-e woodblock print, waves crashing in flat colour with visible wood grain, animated like moving paper cutouts.
Image to video
Every model on this prompt →›What we asked for
The coffee dripper slowly fills as coffee drips through it, steam rising, the camera drifting gently around it, the dripper itself unchanged.
Additional workflow tests
Separate from the common benchmark above. These tests examine longer shots, successive edits or specialized controls. One attempt per test, including failures. Costs are estimates from published rates.
H3 Max Camera Controls: product orbit
View the generated starting image
H3 Max
Camera parallax with a stationary product; this endpoint returns video, not a 3D mesh.
›What we asked for
The coffee dripper is rigid and motionless. Only the camera moves. Preserve its shape and materials and the white background.
minimax/h3-max/camera-controls
prompt_expansion_mode: disabled, camera_trajectory: [{"time":0,"azimuth":0,"elevation":0,"distance":1},{"time":1,"azimuth":45,"elevation":10,"distance":1}], resolution: 768P, duration: 5
returned 768 by 768 pixels
Generated . Cost basis: fal h3-max 768p $0.04/second for 5 requested seconds; 50% launch promotion through September 30, 2026; source read 2026-09-18, not a receipt. Original output.
The product remains recognizable across sampled frames. Its near-symmetric shape makes orbit accuracy difficult to establish from this example alone; a video result is not a reconstructed 3D mesh.
H3 Max Lip Sync: supplied voice
View the generated starting image
H3 Max
A generated portrait plus our synthetic speech sample. This endpoint takes image and audio URLs, not a text prompt. Synchronization has not been scored by ear.
›What we asked for
Animate the generated portrait to the supplied synthetic ad-read audio.
minimax/h3-max/lip-sync/image-to-video
resolution: 480P
returned 832 by 480 pixels
Generated . Cost basis: fal pricing API base rate $0.05/second at the selected 480P, read 2026-09-18; estimated using clipped source-audio duration. Original output.
The generated portrait changes mouth shape while the kitchen and clothing remain recognizable. The output retains supplied audio; synchronization quality has not been scored by ear.
Frequently asked questions
Is Turbo half the price?
At 768p, the published promotion is $0.02/second for Turbo and $0.04 for Max through September 30, 2026. Regular rates are $0.04 and $0.08 respectively.
Did you measure faster-than-real-time inference?
No. Our values measure elapsed request time including the queue and five-second polling interval. They are not raw GPU inference timings.
Voyager
Run any of these inside Voyager
Voyager is an agent for creative work that runs every model on this page. Brief it, and it picks the model for the job, makes the thing and hands back something you can edit. Create an account to get started.