Best AI Models for Video Generation - June 2026

Summary

The leaderboard has reshuffled completely since March. Two models from Alibaba's ATH Innovation Unit - HappyHorse-1.0 and HappyHorse-1.1 - now hold the top two positions on the Artificial Analysis "without audio" rankings at Elo 1,290 and 1,285 respectively. ByteDance's Seedance 2.0 dropped to third in that track but leads the "with audio" category at 1,218, and it's finally accessible globally via fal.ai since April 9, 2026. xAI entered the competitive tier in June with Grok Imagine Video 1.5. OpenAI effectively exited consumer video generation: Sora 2's app closed April 26, the API sunset is September 24, 2026.

Rankings

Elo scores from the Artificial Analysis Text-to-Video Arena as of June 24, 2026. "No Audio" and "With Audio" are separate leaderboard tracks. N/A means the model either doesn't appear in that track or hasn't accumulated enough votes for a stable rating.

Detailed Analysis

HappyHorse-1.0 and 1.1 - Alibaba's Surprise Leadership

The model that now leads all visual quality rankings arrived with almost no announcement. HappyHorse-1.0 appeared on the Artificial Analysis leaderboard around April 7, 2026 with no provider attribution, climbed to the top of blind-test rankings within days, and was later traced to Alibaba's ATH Innovation Unit - a team led by Zhang Di, the former Vice President at Kuaishou who built the original Kling technology. The team joined Alibaba at the end of 2025.

The architecture is a 15-billion-parameter unified transformer with 40 self-attention layers and no cross-attention modules. Joint audio-video generation happens in a single forward pass. The model's strength is in visual coherence rather than audio quality, which is why it leads the "without audio" category (Elo 1,290) while its "with audio" Elo of 1,123 trails Seedance 2.0 by nearly 100 points.

HappyHorse-1.1 adds synchronized multilingual audio with lip-sync across seven languages and bumps output to 1080p. The tradeoff is price: $0.14/s (720p) for v1.0 versus $0.28/s (1080p) for v1.1. A 10-second clip at 1080p in v1.1 costs $2.80 via the fal.ai API.

Clip duration maxes at 5-8 seconds for both versions. There's no multi-shot support within a single generation call. For social content and product demos, those limits are fine. For narrative sequences, you're stitching clips and managing character consistency manually.

Professional video workflows now routinely stitch AI-produced clips together - a limitation of models with short max durations that makes character consistency across clips essential. Source: pexels.com

Dreamina Seedance 2.0 - Now Global, Still Best for Audio

The March verdict on Seedance 2.0 was straightforward: best model, not available globally. That changed April 9, 2026 when fal.ai launched global API access. The "with audio" Elo of 1,218 puts it 66 points ahead of HappyHorse-1.1 in that specific track.

Seedance 2.0 generates audio and video simultaneously using a Dual-Branch Diffusion Transformer. The multimodal input stack - up to 9 images, 3 video clips, and 3 audio files in one call - is more flexible than any other model here. Multi-shot support within a single generation remains a differentiator: Seedance 2.0 can produce coherent scene cuts from a single prompt.

Two things to know before committing to a Seedance integration. First, US regulatory status is unclear: senators pushed back on ByteDance's international deployment in March 2026, and access in the US may be blocked depending on jurisdiction and ISP. Verify current access from your actual deployment region before building. Second, the Pro tier on fal is $0.247/s. The Fast tier at $0.022/s is dramatically cheaper but outputs at lower resolution. Know which tier your production workflow actually needs.

The Seedance 2.0 review covers the difference between Fast and Pro outputs in more detail.

Kling 3.0 - The Reliable Production Choice

Kling 3.0's Elo positions haven't changed much since March - 1,251 (no audio) and 1,104 (with audio). What's changed is competitive context. It used to be the clear global #2; it's now at position 4 in the "without audio" rankings, behind both HappyHorse versions and Seedance 2.0.

Reliability and API stability still matter, though. The Kling API has run at consistent throughput since February with no major reported outages. Pricing at $0.075/s for text-to-video is significantly cheaper than HappyHorse-1.1 ($0.28/s) and Seedance 2.0 Pro ($0.247/s). For teams that need consistent, predictable outputs at production scale, Kling is still the practical choice.

The Kling O3 (Omni) variant at $0.1125/s adds vector-based camera path control and reference-to-video workflows. Elo 1,235 in the "without audio" track reflects stronger video editing performance than the standard 1080p Pro.

Native 4K at 60fps remains a Kling exclusive in the globally available commercial tier. No other provider in this table generates at that resolution and frame rate through a stable API.

See the Kling 3.0 review for frame-by-frame quality comparisons.

Grok Imagine Video 1.5 - xAI Enters the Top Tier

xAI has been running a video generation product since August 2025. The v1.0 entry on the Artificial Analysis "without audio" leaderboard sits at Elo 1,233 - competitive with Kling O3. Version 1.5 moved to general availability on June 16, 2026, making this the most recent major release in the table.

Pricing for v1.5: $0.08/s at 480p, $0.14/s at 720p, $0.25/s at 1080p, with audio included at no surcharge. The model generates 4-15 second clips and xAI claims it ranks #1 on the image-to-video Arena leaderboard.

One important caveat: the Elo 1,233 in the table comes from the v1.0 leaderboard entry. Version 1.5 went live June 16 and hasn't built up enough arena votes for a stable Elo rating at time of writing. The quality improvement from v1.0 to v1.5 is real, but the benchmark backing for v1.5 specifically isn't there yet. Treat vendor performance claims for v1.5 with appropriate skepticism until independent scores arrive.

Veo 3.1 - Price Drop and Audio Quality

Google cut Veo 3.1 Standard from $0.75/s to $0.40/s earlier in 2026. The Fast tier at $0.15/s is now more competitive with Kling 3.0. These aren't trivial cuts - at scale, the pricing difference between Veo and competitors matters significantly.

The Veo 3.1 model page has full technical specs. The Elo score of 1,093 in the "with audio" track reflects solid but not exceptional visual quality. Veo earns its position through audio generation - synchronized dialogue, ambient sound, and environmental audio that's perceptibly cleaner than competing models at similar price points.

Consumer access: Google AI Pro ($19.99/month) and Google AI Ultra ($249.99/month). Developer access: Vertex AI and the Gemini API.

Open-Source Options

The open-source standard in this capability shifted in March. Wan 2.7 from Alibaba replaced Wan 2.6 as the primary self-hostable choice, with meaningful new features.

Wan 2.7, released late March 2026, adds first-and-last-frame generation (specify both endpoints, the model creates the transition), video extension from existing clips, and native audio output at 1080p for up to 15 seconds. The fal.ai API charges $0.10/s, which makes it accessible without owning hardware. In the Artificial Analysis "with audio" leaderboard, it sits at Elo 1,092 - within one point of Veo 3.1.

For teams with hardware, LTX-2.3 from Lightricks (released March 5, 2026) is the open-weights choice for resolution. The 22-billion-parameter model outputs 4K at 50fps using a shared 48-transformer-block architecture across video and audio streams. The rebuilt encoder in v2.3 addresses the soft-output problem from LTX-2 Pro and produces sharper textures at high resolutions. Minimum spec is 12GB VRAM (RTX 3060); comfortable 1080p needs 16GB+.

See the LTX-2.3 model page and the LTX-2.3 review for self-hosting benchmarks.

LTX-2.3 and Wan 2.7 are the two serious self-hosting options in 2026. LTX-2.3 needs 16GB+ VRAM for 1080p comfort; Wan 2.7 needs 8GB minimum. Source: pexels.com

Methodology

All Elo scores come from the Artificial Analysis Text-to-Video Arena. The arena runs blind side-by-side preference voting: evaluators see two outputs from the same prompt without knowing the source model, then vote for the better video. Elo ratings are derived from those comparisons, with confidence intervals based on sample size.

Two parallel leaderboards run: "Without Audio" (all models eligible) and "With Audio" (models that create native audio alongside video). Scores across these tracks aren't directly comparable - a model winning in "With Audio" is completing a harder task.

Caveats that matter for practical use:

• Elo scores shift weekly as votes build up. A 10-20 point difference can flip in days with enough vote volume.
• Clip duration skews human preference in ways that raw Elo doesn't capture. HappyHorse's 5-8 second maximum versus Seedance's 15-second capability isn't reflected in the scores.
• Benchmark contamination is real. Models aware of arena prompt distributions optimize for them.
• Regional availability is less predictable than it was six months ago. Seedance 2.0 may be blocked in the US depending on your provider. Always test access from your actual deployment environment before committing.

Historical Progression

• March 2025 - Kling 2.0 and Runway Gen-3 Alpha were the clear commercial leaders. Open-source lagged commercial quality by a wide margin on motion smoothness.

• June 2025 - HunyuanVideo 1.5 from Tencent became the first open-source model to credibly challenge commercial quality on VBench metrics, scoring 96.4% on visual quality.

• September 2025 - Sora 2 launched with the most consistent physics simulation in any video model at the time. OpenAI removed free consumer access in January 2026.

• October 2025 - Google Veo 3.1 made native audio-video generation a standard feature, shifting expectations for what a commercial video model should include.

• December 2025 - Runway Gen-4.5 launched and topped the Artificial Analysis arena at ~1,247 Elo. LTX-2 from Lightricks became the first open-source model to break Elo 1,100.

• February 2026 - Kling 3.0 released native 4K at 60fps. Seedance 2.0 launched on ByteDance platforms in China and debuted at Elo 1,269.

• March-April 2026 - HappyHorse-1.0 appeared anonymously on the Artificial Analysis leaderboard and climbed to Elo 1,290 within weeks. Alibaba's ATH Innovation Unit was identified as the developer. Seedance 2.0 went globally live via fal.ai on April 9. OpenAI shut down Sora 2's consumer app on April 26.

• June 2026 - xAI released Grok Imagine Video 1.5 (general availability June 16). Alibaba released HappyHorse-1.1 with audio and 1080p output. The top five "without audio" positions are now held by models from Chinese developers (Alibaba-ATH, ByteDance, Kuaishou), with xAI at #6.

The 60-90 day reshuffling cycle noted in March has held. If the pattern continues, the current top positions will face new challengers by September 2026.

FAQ

What's the best AI video generation model right now?

HappyHorse-1.0 leads blind-vote visual quality at Elo 1,290. For audio-native generation, Seedance 2.0 leads at Elo 1,218. Best globally available value: Kling 3.0 at $0.075/s with native 4K.

Which model is best for budget video generation?

Wan 2.7 (Apache 2.0, self-hostable for free). For API use without owning hardware, Seedance 2.0 Fast at $0.022/s is the lowest per-second rate among top-tier models in this table.

Does open-source video generation match commercial quality?

Not at the top yet. LTX-2.3 at Elo ~1,131 and Wan 2.7 at Elo 1,092 trail HappyHorse-1.0 by 160-200 Elo points. The gap is smaller than six months ago and will likely keep narrowing.

Which model produces the best audio with video?

Seedance 2.0 leads the "with audio" Arena at Elo 1,218. Veo 3.1 generates perceptibly cleaner audio quality per independent evaluations. Grok Imagine Video 1.5 includes audio at no surcharge at $0.14/s (720p).

Is Sora 2 still worth using?

The consumer Sora app closed April 26, 2026. The API remains active at $0.10/s (720p) but is scheduled to sunset September 24, 2026. Building new integrations on Sora 2 API now isn't recommended.

How often do video generation rankings change?

Fast - the top position has changed every 60-90 days since mid-2025. Check the lastVerified date on this page and the Artificial Analysis arena directly for current scores.

Related Coverage

• Best AI Video Generators in 2026 - tool comparison with pricing details
• Video Generation Benchmarks Leaderboard - updated rankings by benchmark
• Kling 3.0 Review - hands-on evaluation of the stable production API
• Seedance 2.0 Review - deep dive into the best audio-integrated model
• LTX-2.3 Review - open-source option benchmarked on consumer hardware
• Veo 3.1 Model Page - full specs and pricing breakdown
• LTX-2.3 Model Page - architecture and hardware requirements

Sources:

• Artificial Analysis Text-to-Video Leaderboard
• fal.ai: HappyHorse-1.0 launch
• datanorth.ai: Alibaba HappyHorse reveal
• HappyHorse-1.1 features
• Seedance 2.0 API on fal
• Grok Imagine Video 1.5 release
• Grok Imagine Video 1.5 pricing - TechTimes
• Veo 3.1 pricing breakdown
• Wan 2.7 on fal
• Sora 2 pricing and sunset
• Kling O3 pricing guide
• SkyReels V4 overview
• LTX-2.3 release