Microsoft Foundry adds five AI video models through fal

Microsoft Foundry is rolling out five video-generation models in public preview through a new partnership with fal, the generative-media platform. The lineup, announced in a November 5 blog post and documented in Microsoft Learn, is Kling v3 Pro, LTX-2.5 Fast, LTX-2.5 Pro, MiniMax H3 Max and MiniMax H3 Max Turbo.

All five take text and image prompts, and MiniMax H3 Max also does reference-to-video, using up to 12 reference files across images, video and audio, per the docs. Kling v3 Pro adds native audio generation and custom element support, and its image-to-video mode accepts a start image and an optional end image.

fal calls itself a full stack for generative media, with APIs for more than 1,500 models across image, video, audio and 3D. The deal brings those models into Azure as a first-party service.

The price is the sharpest part of the announcement, and Microsoft's table lists every rate per second of generated video. MiniMax H3 Max Turbo is the cheapest at $0.025 a second for 480p and $0.08 at 1080p. MiniMax H3 Max runs $0.05, $0.08 and $0.16 at the same three resolutions. LTX-2.5 Fast costs $0.09 a second at 720p and up to $0.30 at 4K, and LTX-2.5 Pro runs $0.12 and $0.17. Kling v3 Pro is the most expensive at $0.1232 a second with audio off, $0.1848 with audio on and $0.2156 with voice control.

There is a catch in the fine print. The blog's own text describes the table as pricing "per minute of generated video", while the table's column header reads "Price per video second (USD)". Microsoft says the rates will be posted to the Azure pricing page in the coming weeks, so teams should confirm the billing unit before they budget. Techgenyz reports the same inconsistency and reads the figures as per second.

Kling v3 Pro is the only model on the US Data Zone offering, and Microsoft says Foundry is the first provider to offer a US-hosted Kling endpoint. The four LTX and MiniMax models run on Global Standard pay-as-you-go, and Microsoft says the initial portfolio is hosted in the US. The Learn docs add that fal services run outside Microsoft facilities and that data egresses to fal, where Microsoft's product terms, data protection addendum and certifications do not apply.

fal cofounder Burkay Gur said in the post that the company has "seen incredible adoption this year alone with generative media across all industries from advertising to marketing to gaming."

For creators, the effect is one more route to models that already exist elsewhere. Whether Azure's per-second rates beat buying them from fal directly is a calculation the missing pricing page will settle.

All five models are in public preview now, and no general-availability date has been announced. Microsoft and fal present the partnership at Ignite session BRK256, "Video AI at Scale with fal on Microsoft Foundry", on November 19 at 11:00 AM CT.