What is actually on the shelf
Microsoft Foundry, Azure's platform for assembling AI applications, has not had much to say about video generation. That changed on 5 November, when Microsoft announced that video models from Kling, LTX and MiniMax are rolling out in public preview through a partnership with fal, the generative-media inference company — offered, in Microsoft's words, "as a first-party Azure service".
The shelf has five items. Kling v3 Pro is the cinematic option, with native audio and custom element support. LTX-2.5 comes in two sizes: Fast for drafts, Pro for higher-fidelity output, both with synchronized audio. The MiniMax entries are H3 Max and H3 Max Turbo. All five take text-to-video and image-to-video input, according to Microsoft's announcement post; H3 Max also lists reference-to-video, according to the same post.
First-party service, third-party serving
Read the announcement closely and the structure is more specific than the headline. The models are "served by fal and sourced from the companies that build them". Microsoft supplies the storefront, the Azure billing relationship and the governance tooling. fal runs them.
The MiniMax entries make the layering plainest. H3 Max and H3 Max Turbo are not endpoints operated by MiniMax. They are, in Microsoft's description, "post-trained variants of MiniMax H3" — variants fal tuned itself from H3's open weights, with fal's engineer quoted crediting "fal's post-training and inference optimizations". The Azure entry is fal's edit of MiniMax's model, served by fal, billed by Microsoft. Three different companies are involved in the MiniMax entry — MiniMax built the base model, fal post-trained it, and Microsoft sells it — so a team evaluating 'MiniMax on Azure' should test fal's post-trained variant, not assume it is MiniMax's own endpoint.
Microsoft says the portfolio is hosted in the US and presents Foundry as the first provider to offer a US-hosted Kling 3 endpoint; the desk has not independently verified that 'first' claim from the linked Microsoft post alone, so treat it as Microsoft's claim, not as an established fact. The deployment labels are not uniform, though: Kling is a US Data Zone deployment, the LTX and MiniMax rows are Global Standard, and Microsoft itself advises teams to review each model's data processing and storage terms.
A price table that argues with itself
The announcement publishes a full price table, more than most preview posts do. It also contradicts itself about the unit. The sentence introducing the table says it "shows pricing in USD per minute of generated video". The table's own column header says "Price per video second (USD)". One of those is wrong, and the difference is a factor of 60.
The numbers settle which one. fal sells H3 Max directly at $0.05, $0.08 and $0.16 per second for 480p, 768p and 1080p, with Turbo at exactly half, and Microsoft's table carries the same figures to the cent. Read as per-second prices, the table matches the serving company's own store; read as per-minute, the 1080p H3 Max rate would be about $0.0027 per second ($0.16 ÷ 60), which would be implausibly low for a served video model and is inconsistent with fal's published $0.16 per second for the same model. Budget from the column header, not the sentence. Neither figure is on the Azure pricing page yet — Microsoft says the prices will be listed there "within the coming weeks", so the blog post, unit confusion included, is the only price reference for now.
The table itself rewards a slow read. Turbo is exactly half the price of H3 Max at every resolution, so the faster model is also the cheaper one. LTX-2.5 Fast at 4K, $0.30 a second, is the most expensive line in the table, above Pro's top listed tier of $0.17 at 1080p; the Pro table simply stops at 1080p. Kling is priced by sound rather than resolution: $0.1232 a second silent, $0.1848 with audio, $0.2156 with audio and voice control. Switching from silent ($0.1232/sec) to audio and voice control ($0.2156/sec) raises the rate by $0.0924/sec, a 75% increase ($0.2156 ÷ $0.1232 = 1.75).
In clip terms, at the published rates: a 15-second H3 Max clip costs $0.75 at 480p and $2.40 at 1080p; on Turbo, $0.375 and $1.20. Ten seconds of Kling with voice control comes to $2.16. That is arithmetic from the table, not quoted packages, and it carries the preview caveat: preview prices can move before general availability, for which Microsoft gives no date.
Clip costs at the published per-second rates (desk arithmetic)
- H3 Max, 480p, 15 sec: $0.05/sec, $0.75
- H3 Max, 1080p, 15 sec: $0.16/sec, $2.40
- H3 Max Turbo, 480p, 15 sec: $0.025/sec, $0.375
- H3 Max Turbo, 1080p, 15 sec: $0.08/sec, $1.20
- Kling v3 Pro, voice control, 10 sec: $0.2156/sec, $2.156 → $2.16
What the Azure wrapper buys
Since the H3 prices match fal's own store exactly, the wrapper is not a discount channel. What it buys is consolidation: one Azure contract, one bill, and Microsoft's evaluation and governance tooling around models otherwise reached through a separate fal account. fal's own site offers five free generations a day for signed-in users, at up to 15 seconds each, according to fal's pricing page as read by the desk on 8 November 2026. For first tests, the free route is the vendor's own site; the Azure route starts making sense when procurement, not price, is the obstacle.
Sources
- [1] Microsoft (Azure AI Foundry blog, Tech Community), "Building the next generation of multimodal AI experiences with fal in Microsoft Foundry", 5 November 2026, updated 6 November 2026 (read first-hand in full at Gate 1)Read source
- [2] fal, "MiniMax H3 Max" model and pricing page (information updated 8 September 2026; rates as published 26 August 2026; read first-hand at Gate 1)Read source
- [3] Windows Mode, "Foundry Adds Kling, LTX and MiniMax Video", 8 November 2026 — secondary coverage, used only to establish the story is still moving todayRead source