MiniMax H3, Alibaba Wan 3.0 lead 2026's best AI video models — with an open-weight surprise
A new ranking shows which AI can turn your words or photos into video, how much they cost per second, and why one open model is now the base for most big names.

What changed in a year
The AI video landscape looks very different from twelve months ago. Native audio in generated clips is now expected as standard. Clips commonly run between 15 and 30 seconds. And for the first time, an open-weights model sits inside the top three 1.
Artificial Analysis built its ranking using one main signal: blind human-preference Elo scores from its Video Arena. Evaluators watch two clips made from the same prompt without knowing which model produced each one, then pick the better clip 1. The board used is the AA-Video-T2V v2.1, which includes audio. A separate board measures image-to-video quality .
Kling 3.0 1080p Pro, made by Kuaishou, is fixed at 1000 Elo on both boards as an anchor point. Scores above 1000 mean voters preferred that model over Kling; scores below 1000 mean Kling won more often 12. Most top-ten entries carry a margin of error around ±13 to ±15 points, so ranks three through eight form a close cluster rather than a strict order .
The top five at a glance
Wan 3.0 from Alibaba holds the top text-to-video spot with an Elo of 1175, backed by more than 7,600 human votes 12. It was launched on August 24, 2026, and can produce 30-second 1080p clips from plain text, PDFs, and slide decks . The ability to ingest full documents sets it apart: hand it a presentation and get a structured video back . It costs $0.20 per second, but its open-weights door is closed, and it drops to eleventh place on image-to-video .
Utopai X, built by film studio Utopai Studios in Mountain View, lands just behind at 1172 Elo on text-to-video and leads image-to-video at 1151 12. It was trained by post-training, or refining after release, the MiniMax H3 weights . There is no public API; access runs only through Utopai's PAI platform at 93 credits per second, starting at $15 for 1,000 credits, and developers cannot call it from their own code .
MiniMax H3 comes in third on text-to-video at 1161 Elo and fourth on image-to-video at 1133 12. It became notable as the first open-weights model inside the video top tier . The base model is a 33-billion-parameter single-stream Transformer that uses a 32-billion-parameter Qwen3-VL encoder and outputs video with native 32 kHz stereo sound . MiniMax opened the API on July 31, 2026, and published the weights on August 3, at a price of roughly $0.08 per second for the 768p version . But the license carries real limits: local deployment is blocked in the US, EU, UK, and South Korea, and only the 768p version is open, not higher resolutions . Despite those caps, H3 has become the foundation many others build on .
Dreamina Seedance 2.5 from ByteDance leads image-to-video at 1161 Elo while sitting fourth in text-to-video 12. It rolled out starting July 31, 2026 through Jimeng AI and Doubao Pro, supporting up to 30-second clips with multiple image, video, and audio reference inputs . On Artificial Analysis its normalized price sits at $0.569 per second, though BytePlus lists roughly $0.23 per second for shorter 720p runs, and regional access varies .
Gemini Omni Flash 1.2 from Google is the newest entry on the board, generally available since November 5, 2026 12. It reaches 1152 Elo on text-to-video and 1086 on image-to-video, trading quality for lower cost: about $0.104 per second on Artificial Analysis, or $0.07 per second at 720p according to Google . The model covers 3-to-15-second generations, making it useful when you need quick drafts at budget prices .
The open-weight ripple effect
MiniMax H3's open release triggered a wave of follow-on products. HeyGen Video 1, fal's FLUX 3 Max, Pika's Pika H3 Flex, and Pruna's P-Video-2-Pro all train further on the same base weights 1. Each adjusts the model for a specific use case, from avatar-driven content to social-platform formats, while sharing the same underlying architecture .
Four other models on the list also carry open licenses: FLUX 3 from Black Forest Labs, whose developer promised open weights 1; Pika H3 Flex, restricted open; and LTX-2.5 Fast from Lightricks . Closed-weight leaders fill the rest of the top ten, including SkyReels V4, Dreamina Seedance 2.0, and Google's higher-end Veo 3.1 models, which sit lower on the board at 964 and 948 Elo respectively while costing up to $0.40 per second 2.
How good is it
According to Artificial Analysis's blind human-preference Video Arena (AA-Video-T2V v2.1), MiniMax H3 scored 1161 Elo on text-to-video and 1133 Elo on image-to-video, both measured against an anchor point fixed at 1000 Elo for Kling 3.0 1080p Pro 1. Those scores put it inside the tight top-tier cluster where most ranks three through eight overlap within a margin of error of roughly ±13 to ±15 points 2. On text-to-video it trails only Wan 3.0 and Utopai X, and on image-to-video it trails Dreamina Seedance 2.5 and Utopai X but stays above several frontier rivals like Gemini Omni Flash 1.2 and FLUX 3 Max . Higher Elo means evaluators chose that model more often in head-to-head clips generated from the same prompt, so MiniMax H3's numbers place it firmly among the leading video models of 2026.
Elo score from blind human preference votes, with Kling 3.0 1080p fixed at 1000 as the reference point. Higher is better. Numbers: Artificial Analysis. Chart drawn by Herald of AI.

Picture: artificialanalysis.ai
Key facts
- Wan 3.0 tops text-to-video at 1175 Elo; MiniMax H3 leads as the best open-weights model
- Utopai X ranks near the top on both text-to-video and image-to-video but has no public API
- MiniMax H3 weights are open for download but blocked for local use in the US, EU, UK, and South Korea
- Seedance 2.5 leads image-to-video at 1161 Elo; pricing ranges widely by region and resolution
- Gemini Omni Flash 1.2 entered the rankings on November 5, 2026, offering lower-cost frontier-tier output
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
Apps and productsGoogle opens TorchTPU, a native PyTorch backend for cloud TPUsHerald of AIhttps://github.com/soumith, CC BY-SA 4.0, via Wikimedia Commons
Google opens TorchTPU, a native PyTorch backend for cloud TPUs
A new open-source stack lets developers run PyTorch models directly on Google's TPU chips, replacing the older PyTorch/XLA system.
Apps and productsGoogle Unveils Playground An AI-Powered Gaming Platform Where Anyone Can Create GamesHerald of AI
Google Unveils Playground An AI-Powered Gaming Platform Where Anyone Can Create Games
Google's new experimental platform lets users create, play and share games with simple text prompts—no coding required.
Business and rulesMirendil raises $1 billion to use AI agents for faster model developmentHerald of AI
Mirendil raises $1 billion to use AI agents for faster model development
A new startup founded by former Google DeepMind and Anthropic researchers says its automated platform lets small teams build large AI models quickly and cheaply.