The AI Video Model Releases That Actually Matter (Summer 2026)
Kling 3.0, HappyHorse 1.1, MiniMax H3, Wan 2.7 and Runway Aleph 2.0 all shipped within a few months. What changed in each, what it costs, and which ones are worth switching to.
Five releases between May and July 2026 changed what is possible in creator-facing AI video. Here is what each one actually added, with prices read from the live endpoints on 2 August 2026.
Kling 3.0 (May 2026)
The headline features are 4K output, a Turbo tier and multi-prompt support. The change that matters most for character work is quieter: in Motion Control, character_orientation: "video" raises the reference-clip ceiling to 30 seconds and enables the elements entry that binds a face to the character. In the image orientation, that binding does nothing — a distinction that explains a large share of the "Kling drifts faces" complaints.
Prices: Pro Motion Control $0.168/s, Standard Motion Control $0.126/s, Pro image-to-video $0.140/s, Turbo Standard image-to-video $0.112/s. Native audio is English and Chinese only. fal auto-refunds failed motion-control jobs.
HappyHorse 1.1 (June 2026)
Added native audio, multilingual lip-sync, up to nine character references and 1080p, across nine aspect ratios — including genuine 9:16 and 4:5, which is still unusual and is the reason it is our default for vertical.
What it did not add is a video-edit endpoint. HappyHorse 1.1 ships text-to-video, image-to-video and reference-to-video, and nothing that ingests a source clip. If you were waiting for the upgrade to solve character replacement in an existing video, it does not — that job stays with Kling Motion Control or Wan.
Price: $0.140/s on both i2v and ref2v.
MiniMax H3 (31 July 2026)
2K output, up to nine image references for subject and style, three video clips for motion and three audio clips, all citable by order in the prompt. Durations 5–15 seconds across seven aspect ratios.
Price: $0.26/s on reference-to-video — the most expensive credible option in this class, and released too recently for a meaningful body of user reports. Promising, unproven.
Wan 2.7 edit-video
Underrated and rarely written about. It takes a source video plus a reference_image_url and a prompt, accepts 2–10 second inputs, and bills a flat $0.10/s at both 720p and 1080p — no resolution multiplier, which almost nothing else on this list can say. Setting audio to origin preserves the source soundtrack.
Note that this is a different endpoint from the Wan reference-to-video family, which generates new scenes rather than editing existing ones. The distinction is easy to miss and they behave nothing alike.
Runway Aleph 2.0 (May 2026)
Up to 30 seconds at 1080p in a single shot, with multi-shot handling and keyframe propagation — aimed squarely at the drift and identity problems that appear when you stitch several short generations together. Priced around $0.28 per second in Runway credits.
The practical friction is that it is Runway-only: not on fal, so it means a separate integration, separate billing and separate failure modes. Worth it if long single-shot consistency is your bottleneck; otherwise a lot of plumbing.
What we would switch to
| If your bottleneck is | Move to |
|---|---|
| Face identity in dance clips | Kling 3.0 Pro Motion Control, orientation: video |
| Vertical output that is actually 9:16 | HappyHorse 1.1 ref2v |
| Preserving the original background | Wan 2.7 edit-video |
| Drift across a long single shot | Runway Aleph 2.0 |
| Nothing in particular | stay put, and re-check prices quarterly |
The last row is not a joke. Every endpoint above reprices without announcement, and the release that matters for your workflow is usually not the one with the loudest launch.