fal says H3 Max video model can generate clips faster than real time

fal released benchmark results showing its H3 Max and H3 Max Turbo video models can return 5-second clips with synchronized audio in less time than they take to watch. The company says the models are available now through its serverless API, with pricing starting at 2.5 cents per second during a limited-time discount.

Why it matters: - fal's measurements suggest AI video generation is moving into real-time territory, which could speed up ad creative, product demos and other short-form video workflows. - The benchmarks also show synchronized audio generated in the same pass, reducing the need for separate sound production steps. - Faster-than-playback output at multiple resolutions may make AI video more practical for production use, not just experimentation.

What happened: - fal published measured generation times for H3 Max and H3 Max Turbo, the company's post-trained versions of the open-weights MiniMax H3 video model. - The measurements were taken Sept. 8, 2026 on text-to-video with prompt expansion left at its default setting. - Across nine configurations covering the accepted resolution and clip-length combinations, both endpoints generated video faster than the clips play. - At 768p, the tuned resolution, a 5-second clip generated in 1.54 seconds on H3 Max Turbo and 2.46 seconds on H3 Max. - At 768p, a 10-second clip generated in 4.29 seconds on Turbo and 7.55 seconds on H3 Max. - At 768p, a 15-second clip generated in 8.44 seconds on Turbo and 15.17 seconds on H3 Max. - At 480p, a 5-second clip generated in 0.44 seconds on Turbo and 0.75 seconds on H3 Max. - At 1080p, a 5-second clip generated in 2.33 seconds on Turbo and 3.12 seconds on H3 Max.

The details: - The reported figures are generation time, meaning the inference value each endpoint returns in its own response. - Caller wait time can be longer because queue time, prompt expansion and encoding come on top of generation time. - Turbo returned all nine configurations faster than real time. - H3 Max returned seven of nine configurations faster than real time. - A default request returns a 5-second, 768p clip at 1344 by 768 pixels and 24 frames per second, with stereo audio in the same file. - Dialogue, room tone, effects and music are described in the prompt rather than added in a second step. - Picture and audio are generated in the same pass. - fal says the speed comes from designing the model and the inference engine as one system rather than training first and serving later. - fal Research added new training data and used an in-house reinforcement learning framework to improve prompt adherence and visual quality. - fal kept optimizations such as lower precision and reduced sampling steps only where human preference scores held. - Artificial Analysis ranks H3 Max first among image-to-video models with audio output in its Image to Video Arena, with an Elo score of 1200. - The next model in that ranking scores 1192, and the base MiniMax H3 model scores 1187.

Between the lines: - The benchmark pitch is not just about speed. It is also about making audio and video arrive together, which can simplify content creation pipelines. - The combination of serverless delivery, long prompts up to 50,000 characters and multiple aspect ratios points to a product aimed at developers and production teams, not only hobbyists. - The claim of being faster than playback matters most when paired with the lower-cost Turbo tier, since speed and price often decide whether teams adopt a model in practice.

What's next: - H3 Max and H3 Max Turbo are available now through fal's serverless API. - The API offers six endpoints: text-to-video, image-to-video and reference-to-video on H3 Max, text-to-video and image-to-video on H3 Max Turbo, and H3 Max Director, a realtime model that holds a streaming session instead of answering a single request. - Clips run from 5 to 15 seconds at 480p, 768p or 1080p, in aspect ratios from 21:9 to 9:16. - Python and JavaScript SDKs are available, and the API can also be called over HTTP. - Pricing is per second of output video with no minimums and no subscription. - H3 Max lists at $0.05 per second at 480p, $0.08 at 768p and $0.16 at 1080p. - H3 Max Turbo lists at $0.025, $0.04 and $0.08. - Both products are 75% off through Sept. 14, 2026. - Signed-in users get five free generations a day of up to 15 seconds each on a rolling 24-hour reset. - H3 Max and H3 Max Turbo are available at fal.ai.

The bottom line: - fal is betting that faster-than-real-time AI video generation, paired with synchronized audio and usage-based pricing, can move video creation from a novelty to an everyday API workflow.