ai& video is a hosted service for generating H3 clips through a browser or API, with processing on the company's infrastructure in Japan. Its launch distinguishes a step-distilled Fast tier from a Quality tier with broader reference inputs. The useful choice is the actual request contract and quote, rather than assuming every H3-branded option has the same inputs or speed.
ai& announced the service on October 22, 2026. This guide checks its engineering explanation, Videos API, and public video-model catalog on October 23. We made a read-only catalog request, not a generation request. H3Vid has not integrated ai& or reproduced its hardware benchmarks.
Compare the advertised tiers
The launch engineering page describes Fast as step-distilled H3 running on Tenstorrent, with text and first/last-frame inputs. Quality runs on AMD using Cache-DiT, with an additional reduced-precision attention recipe for requests longer than sixteen seconds. It also supports multimodal reference generation.
| Launch tier and input | Published 768p USD rate | 10-second arithmetic example |
|---|---|---|
| H3 Fast, text or frames | $0.0125 per second | $0.125 |
| H3 Quality, text or frames | $0.055 per second | $0.55 |
| H3 Quality, references | $0.075 per second | $0.75 |
These are launch-page rates, not observed invoices. USD and JPY are separate list prices on that page; do not treat one as a live currency conversion of the other.
The unauthenticated catalog returned one model, minimaxai/minimax-h3, at $0.055 per second for 768p when we checked. It did not expose a separate Fast model ID. The generic documentation's sample catalog still shows $0.08, which differs from that live response. Discover the current models and inspect the submitted job's quote instead of copying an example price or inventing a Fast endpoint name.
Complete the account agreement before automation
The Videos API requires the individual key owner to accept the current Video Service Terms in the console. An API key cannot record that agreement, and a teammate's acceptance does not cover everyone in the organization. A missing agreement returns 403 agreement_required even when listing models succeeds.
After that setup, the documented pattern is to create a job, poll its status, and download its MP4. The create request includes a catalog model ID, a prompt, requested seconds, and an aspect ratio. The current contract lists 4–30 seconds, with 16–30 marked experimental, and describes 768p output with audio.
Japanese scene descriptions are translated to English before generation. The documentation says dialogue inside its dialogue tags and depicted text are left unchanged. If precise wording matters, keep that wording distinct from the surrounding scene brief and verify the saved soundtrack and visible text afterward.
Choose frames or semantic references
For an opening or ending composition, upload one or two images and pass their file IDs with unique first-frame or last-frame roles. A single ending image is allowed. These frame inputs cannot be combined with the separate multimodal references object.
Reference generation permits up to nine images, three videos, and three audio clips, with twelve files total. Audio needs at least one image or video anchor. Inputs also share a 256 MiB aggregate limit; a file used twice counts twice.
The per-file generation limits are 30 MiB for images, 50 MiB for video, and 15 MiB for audio. They differ from the Files API's general upload ceiling. Use the correct upload purpose and format, then retain the returned file ID. A successful generic upload does not establish that the same file fits a video slot.
For a first reference trial, use one approved picture and a short, clearly relevant clip. References longer than the requested output are trimmed to their first requested seconds. Prepare the desired excerpt beforehand if the useful action begins later in the source. The documentation recommends short reference clips of two to fifteen seconds.
Follow the job through completion
The job can pass through moderation, translation, queueing, and active rendering before completion. The API guide recommends polling every ten to thirty seconds. Do not interpret an unfamiliar intermediate status as a failed render; the guide allows new running stages to be added.
Download the content only after completion. That endpoint returns raw MP4 bytes rather than another JSON object. A thumbnail may also be available, but its presence does not replace checking the actual clip's dimensions, motion, duration, and sound.
The launch's shortest-time claim and the API guide's expectation of multi-minute jobs measure different practical situations. Engine-only hardware tables also exclude MP4 encoding and transport. None establishes a fixed end-to-end latency for your account or request.
Budget, preserve, and evaluate the result
The API fixes cost and currency at submission from the quoted rate and requested duration. Admission requires enough balance for the new quote plus jobs already in flight. The documented balance deduction occurs only on successful completion; failures and timeouts are not charged. Submitted jobs cannot currently be canceled.
Finished jobs and content expire thirty days after completion. Input uploads have their own expiry schedule, and files with less than an hour remaining are rejected for generation. Download approved outputs and preserve your prompt, selected model, source files, and quote outside the provider's temporary storage.
Judge a usable shot by continuity, requested action, and intelligible audio rather than a speed number alone. For another provider-specific execution contract, compare Pruna Reference. For H3Vid's separately implemented online reference tool, see MiniMax H3 Ref2VA. Neither is a substitute identity for ai& Fast or Quality.