Microsoft Foundry Adds Five fal Video Models: Pricing Unit Confusion and Data Residency Caveats

What Microsoft announced​

The Microsoft Foundry blog post is dated November 5, 2026, and was updated November 6. The models give developers different starting points for generating video from text and for animating existing images. Microsoft describes the offering as a first-party Azure service. The models are served by fal and sourced from the companies that build them, so Microsoft did not build them.

One analysis notes that Foundry customers get a curated set of five fal-served models. They do not get fal's full catalog of more than 1,500 models. Microsoft says more models are planned. It does not say which ones or when.

The five models​

Microsoft's descriptions, as relayed by the source reporting, break down like this:

• Kling v3 Pro: Microsoft describes cinematic visuals, fluid motion, native audio generation and custom-element support. Its pricing table lists text-to-video and image-to-video. Microsoft's prose also mentions detailed reference-to-video control. The table does not list reference-to-video as a separate Kling mode.
• LTX-2.5 Fast: This is the speed-optimized option for rapid iteration and previews. It supports text-to-video and image-to-video with synchronized audio. Resolutions run from 720p to 4K.
• LTX-2.5 Pro: This is the quality-optimized option for final output. It supports text-to-video and image-to-video with synchronized audio, at 720p and 1080p.
• MiniMax H3 Max: The table lists text-to-video, image-to-video and reference-to-video at 480p, 768p and 1080p. Microsoft calls the H3 models post-trained variants focused on prompt adherence, aesthetics and inference performance.
• MiniMax H3 Max Turbo: The table lists text-to-video and image-to-video at the same three resolutions. It does not list reference-to-video.

Microsoft offers no benchmarks or measured speed comparisons. The Fast/Pro and Turbo labels describe intent, not proven results.

Listed pricing​

Microsoft's table labels the rates "price per video second (USD)".

The unit discrepancy: The sentence above Microsoft's table says it shows pricing per minute of generated video for LTX and MiniMax. The table's own column header says per video second. The figures above follow the table. Treat the unit as unconfirmed until it is settled. The difference is a factor of 60, so it matters.

Microsoft says the prices will be listed on the Azure pricing page "within the coming weeks". That is a promise, not a sign the listing has appeared. Check the Azure pricing page and your own billing before forecasting spend.

A rough illustration, assuming the per-second reading is right: a 10-second 1080p clip would cost about $1.70 on LTX-2.5 Pro and about $0.80 on MiniMax H3 Max Turbo. Those numbers come from simple multiplication of the table rates. They are not Microsoft estimates.

Data residency: not all five are the same​

Microsoft says the initial portfolio of video models is hosted in the US. The deployment labels differ, though:

• Kling v3 Pro is labeled "US Data Zone".
• LTX-2.5 Fast, LTX-2.5 Pro, MiniMax H3 Max and MiniMax H3 Max Turbo are labeled "Global Standard".

Microsoft's own guidance is to weigh hosting location as one factor. It also tells customers to review each model's documented data-processing and storage terms. A "US-hosted" headline does not mean all five models carry the same residency guarantee. Compliance teams should check each model's terms separately.

Preview caveats​

• The models are in public preview. Microsoft gives no general-availability date.
• Public preview means capabilities, limits, availability and pricing may change.
• Microsoft says teams can use Foundry's evaluation, governance, security and observability capabilities "supported for their selected offering". That wording does not promise every control for every model.

A practical example: ComfyUI​

Microsoft has also published a companion walkthrough. It brings the fal MiniMax H3 Max video model and Foundry image models into ComfyUI. It uses two custom nodes and keyless Microsoft Entra ID authentication. The walkthrough covers deploying the model in a Foundry project, including through the Azure CLI. It notes that fal jobs are asynchronous, so the node polls until the video is ready, then saves the MP4. Build job handling for that polling behavior into any app you write.

How to approach evaluation​

1. Start with the input. Use text-to-video for open-ended concepts and image-to-video to animate an approved image. Choose MiniMax H3 Max if you need reference-to-video, since it is the only model with that mode in the pricing table.
2. Draft cheap, finalize carefully. Microsoft positions LTX-2.5 Fast for previews and LTX-2.5 Pro for final output. MiniMax Turbo is the lowest-priced listing.
3. Price audio explicitly. Kling's rate changes with audio and voice-control settings. Resolution tiers drive the LTX and MiniMax rates.
4. Settle the billing unit. Confirm per-second or per-minute before any budget goes to finance.
5. Document residency per model. Record the deployment type and processing terms for each model you test.

Analysis​

The strongest argument for the move is convenience. Teams already building on Azure can test several video models behind one platform and one set of identity and governance tools. That is a real benefit for enterprises that don't want to onboard each model vendor separately.

The limits are just as clear. The performance claims here come from Microsoft and fal, not independent testing. fal's own site says H3 Max ranks first on quality against leading video models. That is vendor marketing, not a third-party benchmark. Preview terms, an unconfirmed billing unit and mixed residency labels all argue for a small pilot before any production commitment.

The bottom line is that Foundry has more video options to test. Pricing, residency and production readiness still have to be verified model by model.