April 2026 has been a disorienting month for anyone building on AI video.
OpenAI shut down Sora on March 24. A model called HappyHorse-1.0 appeared from nowhere, topped every benchmark by the widest margin in leaderboard history, then turned out to be Alibaba. ByteDance’s Seedance 2.0 rolled out globally — except the United States, where it stayed offline due to copyright disputes with Hollywood studios. The model that sits at #1 today could be gone, blocked, or overtaken by next week.
If you’re building video workflows in 2026, betting your production pipeline on a single AI video model is a structural risk.
The Landscape Shifts Faster Than Any Single Bet Can Track
Three things happened this month that illustrate the point:
Sora is gone. OpenAI’s flagship video model generated $2.1 million in lifetime revenue against $15 million per day in inference costs. The unit economics never worked, and it shut down without a clear replacement timeline. Creators who had built workflows around it had to find alternatives overnight.
A new #1 appeared in days. HappyHorse-1.0 appeared on the Artificial Analysis video leaderboard on April 7, unattributed. Within days it held a 107-point Elo lead over second place — the largest gap in leaderboard history. It turned out to be from Alibaba’s ATH AI Innovation Unit. No one saw it coming.
Legal and geopolitical risk is real. ByteDance’s Seedance 2.0 has been blocked from US distribution due to copyright disputes with Hollywood studios. Geopolitical and legal exposure is now a concrete factor in model availability, not a theoretical one.
Each Model Is Good at Different Things
Even setting availability risks aside, no single model covers all use cases well. The current production-ready landscape roughly shakes out like this:
- Runway Gen 4.5: Leads the Artificial Analysis text-to-video benchmark. Strong cinematic realism, reliable physics — the best choice for editorial-quality clips.
- LTX Video: Optimized for speed and determinism. Supports first and last frame control, useful when you need to generate a clip that bridges exactly between two reference frames.
- Kling 3.0: Excels at multi-shot narrative with per-shot camera and pacing control. 4K output and native lip-synced audio in one pipeline.
- Hailuo (MiniMax): Strong cinematic portrait video. The go-to for facial micro-expression realism and close-up talking-head content.
Trying to do all of this with one tool means accepting compromises on most of it.
What a Multi-Model Workflow Actually Looks Like
In practice, routing generation tasks to the right model looks like this:
- Use Runway Gen 4.5 for wide establishing shots and cinematic b-roll.
- Use LTX with first/last frame control to generate transition clips that match your existing cut.
- Use Kling for structured multi-shot sequences where per-shot camera control matters.
- Use Hailuo for close-up portrait or talking-head content where facial detail is critical.
The overhead is tool-switching and format juggling — unless you’re working in an environment that puts all of them in the same interface.
How Tellers Fits In
Tellers integrates Runway Gen 4.5, LTX Video, and Kling directly in the same editor and agent workflow. You can generate clips from different models, preview and compare them, and assemble the results in a single project — without switching tools or converting formats.
From a single prompt or agent instruction, you can specify which generation model to use for each task. Tellers handles the underlying API calls, timeline placement, and output management.
HappyHorse-1.0 is not yet on Tellers — it’s still in beta and not publicly available through any API. When production access opens up, it’s the kind of model that would fit naturally into the platform. Until then, the integrated lineup of Runway, LTX, and Kling covers the main production use cases well.
FAQ
Which AI video model is best in 2026?
No single model leads across all dimensions. Runway Gen 4.5 leads on text-to-video quality benchmarks, Kling 3.0 on multi-shot narrative control, and Hailuo on portrait realism. HappyHorse-1.0 tops overall leaderboards but is not yet in production.
Is Sora still available?
OpenAI discontinued the original Sora consumer app on March 24, 2026. Sora 2 is available to OpenAI Plus and Pro subscribers via ChatGPT.
Does Tellers support multiple AI video generation models?
Yes. Tellers integrates Runway Gen 4.5, LTX Video, and Kling in a single interface and agent workflow.
What is first and last frame control in video generation?
It lets you specify both the starting and ending frames of a generated clip. The model produces a video that begins at your reference image and ends at another, making it useful for generating scene transitions that fit precisely between existing footage. Tellers supports this for LTX Video.
Why did Sora shut down?
OpenAI cited unsustainable inference costs — approximately $15 million per day — against limited commercial traction. The shutdown underscores why relying on any single AI video platform carries real risk.