피파 한 줄 정리: Video landscape: Sora가 #1 → bankrupt 됐어 (2026-03). Runway Gen-4.5·Veo 3.1·Kling 3.0·Hailuo 2.3·LTX-2·Wan 2.6. Production은 항상 fallback 필요.
If image models are like cameras, video models are like entire film crews — each with different specializations, budgets, and creative philosophies. Some prioritize visual fidelity. Others nail motion physics. Some generate sound alongside the visuals. And one high-profile crew just went bankrupt and left the set.
The Sora Story: A Cautionary Tale
OpenAI's Sora launched to enormous hype, hit #1 on the App Store in September 2025, and then shut down on March 25, 2026. The reason? It cost an estimated $15 million per day to operate against just $2.1 million in total revenue. Even Disney's $1 billion investment deal was canceled. The lesson: a technically impressive model that can't sustain itself economically isn't a tool you can build workflows around.
Runway Gen-4.5
As of early 2026, Runway Gen-4.5 leads quality benchmarks at the top of the Elo rankings. It generates up to 60 seconds of 1080p video with native audio generation, excels at physics-accurate motion and character consistency, and costs roughly $0.20–$0.50 per second. Runway has the most mature editing and post-production ecosystem built around its generation capabilities.
Google Veo 3.1
Veo 3.1 offers native 4K output at 60+ seconds with synchronized audio generated in a single pass. For professional filmmaking-oriented use cases — where resolution, color science, and cinematic feel matter — Veo currently leads. Its tight integration with Google's infrastructure makes it particularly attractive for enterprise and studio workflows.
Kling 3.0 (Kuaishou)
Kling 3.0 is the value leader: native 4K at approximately $0.50 per clip or $6.99/month, with multi-shot storyboarding support. The tradeoff? Maximum clip length is 15 seconds. If your workflow involves short, punchy clips edited together (which is often the smartest approach anyway), Kling delivers remarkable quality per dollar.
MiniMax Hailuo 2.3
Hailuo excels at physics-based motion and complex instruction following. Its Start and End Frame feature gives you precise control over motion trajectories. It supports diverse art styles from photorealism to anime to ink wash painting, making it versatile for stylized work. The 2.3 Fast variant cuts costs by 50% for batch exploration.
Open-Source Video: LTX-2 and Wan 2.6
LTX-2 (Lightricks) is open-source at 4K resolution. Wan 2.6 (Alibaba) is completely free and open at 1080p. Both are viable for local deployment and experimentation, though they trail proprietary models in motion quality and consistency.
- The video model landscape moves fast — Sora went from #1 app to shutdown in six months.
- Quality, cost, duration, resolution, audio, and controllability are all independent axes. No model leads on all of them.
- The smartest workflow often uses cheap fast models for exploration and premium models for hero shots.