State of AI Video — April 2026

The AI video leaderboard just had its biggest shakeup since Runway Gen-4.5 debuted in December. HappyHorse 1.0 became a major watchlist model because of reported benchmark strength and multilingual lip-sync, but we are not treating it as a verified self-hostable release yet. Meanwhile, 6 major model launches since February made Q1 2026 the busiest quarter in AI video history.

This is the first issue of VidScore’s monthly State of AI Video report. Each month we’ll cover Arena rankings, new launches, pricing shifts, and the trends shaping this market.

Data as of April 29, 2026. ELO scores shift daily; check the live leaderboard for real-time rankings.

Key Numbers This Month

• New API candidate: HappyHorse 1.0— public API and generation pricing verified; public weights not verified; check the availability tracker
• 6 major launches in Q1 2026: Kling 3.0 Omni, Seedance 2.0, PixVerse V6, Wan 2.7, SkyReels V4, and Veo 3.1 Lite
• 28 model records now tracked on VidScore with verified API pricing
• $0.02–$0.70/sec API pricing range — a 35x spread
• Sora shutdown: consumer app/web ended April 26; API ends September 24
• 3 of the top 5 Arena models are open source or open-weight

Arena Rankings: April 2026

The Artificial Analysis Video Arena ranks models via blind A/B comparisons where humans pick the better video. Here are the current top 10 for text-to-video quality (without audio track):

Pricing shows lowest available API tier (provider trust tiers vary — see individual provider guidesfor details). ELO scores from the “without audio” T2V track; rankings differ in the audio-inclusive track.

Notable: Runway Gen-4.5peaked at ELO 1,247 and #1 when it launched in December 2025. Four months later, it sits at #7 in the table above, behind both Kling 3.0 and Grok Imagine Video — models that cost 2–5x less. The pace of displacement is accelerating: no model has held #1 for more than 8 weeks in 2026.

What Changed This Month

HappyHorse 1.0: API-Available, Not Verified Self-Hostable Yet

The biggest story this month is HappyHorse 1.0from ATH-AI, the team within Alibaba’s Taotian Group led by former Kuaishou VP Zhang Di. It is being discussed as a major quality contender and now has FAL and Replicate provider APIs, but we are not publishing it as a verified open-source/self-hostable release until public weights, inference code, and license terms are confirmed.

What makes this worth watching: public reporting attributes strong quality, fast H100 inference, and multilingual lip-sync to the model. What is now verified for builders: provider API access and generation pricing. What we cannot verify yet: downloadable weights, runnable inference code, self-hosting license terms, and a mature official public repo.

Until those release artifacts are public, treat HappyHorse as an API candidate rather than a self-hosting target. For a full breakdown, see our HappyHorse 1.0 review and the HappyHorse availability tracker.

Kling 3.0 Omni: Kuaishou Goes Full Multimodal

Kling 3.0and Kling 3.0 Omni launched February 5 as Kuaishou’s most ambitious release. The Omni variant supports full multimodal I/O (text, images, audio, video simultaneously), multi-shot storyboarding with per-shot camera control, and native dialogue across multiple languages. At $0.112/sec (Pro tier on FAL.ai without audio), it enters the Arena at ELO 1,243in text-to-video — slotting in just behind SkyReels V4 and ahead of Grok Imagine Video. For details on features and pricing, see our Kling v3 review.

Veo 3.1 Lite: Google Cuts Prices in Half

Google launched Veo 3.1 Lite on March 31, its most cost-effective video model. At $0.05/sec (720p) and $0.08/sec (1080p), it costs less than half of Veo 3.1 Fast while maintaining the same generation speed. The same week, Google cut Veo 3.1 Fast pricing to $0.10/sec (720p) and $0.12/sec (1080p). Both models support text-to-video and image-to-video with native audio. For the full Veo lineup breakdown, see our Veo 3 review.

Seedance 2.0: Multimodal Power, Copyright Trouble

Seedance 2.0launched in late February as ByteDance’s most ambitious video model. Its defining feature: a unified architecture that accepts text, up to 9 reference images, 3 reference videos, and 3 audio files simultaneously. No other model accepts this many input types. At ELO 1,273 (T2V without audio), it ranks #2 behind HappyHorse.

However, in mid-March ByteDance paused the global rollout after the Motion Picture Association, Disney, Warner Bros., Netflix, Sony, and Paramount sent cease-and-desist letters alleging the model was trained on copyrighted material. ByteDance pledged to add safeguards and stopped generating realistic human faces and IP-protected characters. The model remains available via 7+ API providers at $0.10–$0.30/sec verified (FAL.ai at $0.303/sec is the highest-trust option), but the copyright cloud raises questions about long-term availability. See our Seedance 2.0 review for a full assessment.

PixVerse V6: Cinematographer-Grade Controls

PixVerse V6 (March 30) entered the Arena at #16 with ELO 1,209 for text-to-video, but jumped to #4 for image-to-video(ELO 1,313). Its standout feature: 20+ cinematic lens controls including focal length, aperture, depth of field, and chromatic aberration — all adjustable via prompts.

Pricing is aggressive: $0.025/sec at 360p up to $0.115/sec at 1080p with audio. The widest resolution-based pricing ladder of any model.

Wan 2.7: Four Modes, One Architecture

Alibaba’s Wan 2.7 shipped in March as a hosted/API video model with a 27B-parameter Mixture-of-Experts architecture. It consolidates four generation modes into one architecture: text-to-video, image-to-video with first/last-frame control, reference-to-video with voice cloning, and instruction-based video editing.

At $0.10/sec on FAL.ai, it offers a wide feature set per dollar. We do not currently verify downloadable Wan 2.7 weights, runnable inference code, or open-source licensing for self-hosting.

Winners and Losers

Winners

• Open-weight workflows. Verified self-hostable models like FramePack, HunyuanVideo, and LTX-2 Pro are now practical enough for real workflows. HappyHorse and Wan 2.7 are high-interest watchlist items, but they should not be counted as verified self-hostable releases yet.
• Native audio. Every major launch in 2026 includes audio generation. Silent-only models (Pika, Hailuo, CogVideoX) are increasingly at a feature disadvantage. Joint audio-visual generation is the new baseline.
• Budget creators.Google’s Veo 3.1 Lite at $0.05/sec joins Grok Imagine Video ($0.05/sec) and Kling 2.5 Turbo ($0.042/sec) in the budget tier. Three competitive models now cost under $0.06/sec.
• Chinese AI labs. 4 of the top 5 Arena models are from Chinese teams: ATH-AI (HappyHorse), Skywork AI (SkyReels), Kuaishou (Kling 3.0), and Shengshu (Vidu).

Losers

• Runway’s lead.Gen-4.5 held #1 for about 8 weeks after its December launch. It’s now #7 in the T2V table with six models ahead of it — five of which cost less. At $0.12/sec, it’s competitively priced among top-tier models. Runway’s $10M startup fund (launched March 31) and $315M Series E at $5.3B valuation show the company is betting on ecosystem, not just model quality.
• Sora’s final chapter. OpenAI confirmed the Sora consumer app shuts down April 26, 2026, with the API following September 24. The closure stems from high compute costs, declining users (from 1M to under 500K), and mounting copyright lawsuits from Hollywood studios. For alternatives, see our Sora 2 review.
• Copyright-entangled models.Both Sora and Seedance 2.0 face copyright-related setbacks. The MPA and individual studios have drawn a line. Models that can’t prove clean training data face legal and distribution risks.
• Audio-free models.Models without native audio — Pika 2.0/2.5, Minimax Hailuo, Luma Ray 2, CogVideoX — are missing an increasingly mandatory feature. Post-production audio adds cost and friction that competitors eliminate.

What’s Coming

• HappyHorse availability: Hosted API access and generation pricing are now verified. Watch next for public weights, inference code, license terms, and provider reliability data before moving it into self-hosting guidance. Use the HappyHorse availability tracker and compare HappyHorse vs Seedance before waiting on it.
• Wan 2.7 open-weight status: Track whether Alibaba publishes downloadable weights and self-hosting terms. Until then, Wan 2.7 remains a hosted/API model in this catalog.
• Runway response: With Gen-4.5 sliding from #1 to #7 in four months, expect Runway to accelerate its next release. The $315M Series E and $10M startup fund buy time, but competitive pressure is mounting.
• Sora sunset fallout:Sora’s consumer app closes April 26. Users migrating to alternatives will drive a wave of comparative testing. Expect Arena vote volumes to spike.
• Seedance 2.0 copyright resolution:ByteDance’s response to Hollywood’s cease-and-desist letters will set a precedent for the industry. If safeguards satisfy studios, the global rollout resumes; if not, competitors benefit.
• Price compression continues: Veo 3.1 Lite at $0.05/sec joins the sub-$0.06 tier. The floor has dropped from ~$0.05/sec a year ago to $0.02/sec today. Expect sub-$0.01/sec options by Q3 as open-source self-hosting scales.

Next month’s report will cover HappyHorse availability, Wan 2.7 open-weight status, Sora sunset migration patterns, and any new Arena entries. Follow our live leaderboard for real-time ranking updates between reports. For pricing across all models mentioned, see the AI Video Pricing Guide 2026. For detailed reviews of the top models: HappyHorse 1.0 Review, Kling v3 Review, Seedance 2.0 Review, Veo 3 Review. For open-source options, see our open-source AI video guide.

FAQ