MiniMax H3 Max Renders Video Faster Than It Plays. That Is How AI Employees Get a Face.

§ 01What shipped#

MiniMax launched H3 on July 31, 2026 as a general-purpose multimodal generation model: video with native stereo audio, up to 15 seconds at 2K, with a promise to open the weights. fal took those weights, post-trained them with new data aimed at prompt adherence and visual quality, and built an inference engine around the result. fal’s announcement is careful about what that means: “Traditional step-distillation methods aim to match their base model while we aim for a better model at much faster speed.”

§ 02The speed, with the receipt#

fal’s exact wording: “It does so while generating a 5-second video under 3 seconds of wall time, which is roughly 35x the throughput of the official MiniMax H3 endpoint and on average 15x faster than anything with comparable quality.” The product page adds the mechanism for checking it: every API response carries a timings.inference field, the actual denoising time on the backend, “which lands at roughly 2.5 seconds for a 5-second 768p generation.” fal’s published example response on the image-to-video endpoint shows 2.77 seconds.

That is the claim, and it is a testable one, which is more than most video launches offer. Two honest edges to it. The wall time a user sees includes upload, queueing, and download, so “under 3 seconds” describes the model, not your app’s round trip. And the ratio holds at 5 seconds; at 15 seconds fal says the clip takes about 15 seconds, so the model runs at roughly realtime on the longest setting and roughly twice realtime on the shortest.

§ 03The quality, by people who are not fal#

fal ran its own head-to-head preference study against twelve models, including the official MiniMax H3 endpoint, Gemini Omni Flash, Wan 3.0, Seedance 2.5, Kling 3, and Veo 3.1, scoring overall preference, prompt understanding, and aesthetics with Bayesian Elo ratings and 95% confidence intervals. H3 Max ranks first on all three. That is a vendor study, and we treat it as one.

The independent boards say the same thing, which is why the speed number is worth a page.

The Design Arena gap over the base model is eight Elo points: a post-train that is faster and slightly better, rather than faster and worse. That is the part fal’s announcement calls moving the frontier, and on these two boards it is a fair description.

§ 04The price, and the cost table fal wants you to read#

Standard rates match on both APIs. MiniMax’s pay-as-you-go page lists MiniMax-H3-Max at $0.05 a second for 480p and $0.08 for 768p, billed on output video only; standard H3 is $0.08 at 768p and $0.13 at 2K. fal’s launch discount, 75% off on its model pages as of this writing, ends September 7.

Same price as the standard model at 768p, about 35 times the throughput. The trade is 2K and reference-to-video, which stay on standard H3.

fal also publishes a cross-model cost table for 15 seconds of roughly 720p video at fal’s own August 26 rates, with Kling quoted on its cheapest v3 tier that generates audio, since H3 Max always does. It is a vendor’s table on a vendor’s platform, and prices elsewhere differ, but the ordering is the point.

§ 05The line that matters: generation faster than playback#

Here is why a video model earns a page on a blog about AI employees.

Every AI employee today is text. It writes the report, files the task, sends the email, refreshes the dashboard. Video, when it appears at all, is a production step: a script, a render that takes minutes, a review, a publish. That shape means video is something an employee makes occasionally, the way a human employee occasionally records a screencast. It is not something an employee is.

Faster-than-realtime generation changes the shape. When a 5-second clip renders in under 3 seconds, a video reply costs about what a paragraph costs in wall time. An employee with a reference face and a cloned voice can answer a question with a short talking clip, in the thread, while the question is still warm. The owner’s morning status can be a face saying it. A customer’s how-do-I question can get a fifteen-second demonstration instead of a wall of steps. None of that is a marketing video. It is the ordinary work of a standing role, now with a face on it.

Three properties have to hold at once for that to be real, and H3 Max is the first model to publish credible numbers on all three:

• Speed past realtime. Under 3 seconds for 5 seconds of video, with the timing exposed in the API response.
• Quality good enough to represent someone. First on two independent image-to-video boards; a face that drifts or a mouth that fails to sync is worse than text.
• A price that survives frequency. $0.40 for a 5-second reply at 768p, so a worker that answers in video a dozen times a shift is spending single-digit dollars, not a production budget.

What is missing is identity control. MiniMax’s API reference is explicit that MiniMax-H3-Max does not support reference-to-video: you cannot hand it a reference image, video, or voice clip to lock a character. Image-to-video with a first frame is the workaround, and fal’s own examples show one character holding across six locations in a single generation, so the model can hold a face. But a standing employee needs its face locked across thousands of generations over months, not six shots in one, and that is the reference-to-video job. When it lands on the Max variant, or on a competitor at the same speed, the third property is met and the unlock is complete.

§ 06What this means for CellCog#

CellCog’s video generation runs on Seedance 2.5 today, chosen because it accepts dozens of reference images, videos, and audio clips per segment and assembles multi-minute films from one prompt: the production job, done well. CellCog avatars already carry the other half of the pairing, a reference face and a cloned voice, and every avatar video ships disclosed as AI-made.

H3 Max is the first model whose numbers make the third thing plausible: a face that answers at conversational speed. It is not there yet, for the reason above, and neither is any competitor at this speed. So the honest position is a watch, not a switch. The signals this page updates on: reference-to-video arriving on H3 Max or on any faster-than-realtime model, independent replication of the under-3-seconds figure outside fal’s own timing field, and the September 7 end of fal’s discount. If a model clears the bar, CellCog employees get a face the same way they got Gemini 3.8 Flash and Fable 5.1: underneath the role, on the day it is ready, with nothing to change on your side. A full shift of real work runs about $25, and the cost depends purely on how much work you assign.

§ 07The record#

As of September 3, 2026: H3 Max is available on fal (text-to-video and image-to-video endpoints, free sandbox at five a day) and on MiniMax’s pay-as-you-go API as MiniMax-H3-Max; list price $0.05/s at 480p and $0.08/s at 768p; fal’s launch discount ends September 7; no 2K and no reference-to-video on the Max variant; #1 on Artificial Analysis image-to-video with audio (1,201) and Design Arena image-to-video (1,341). If reference-to-video ships on H3 Max, if a competitor publishes a comparable faster-than-realtime figure with comparable rankings, or if the prices move, the update happens here, same day.