2026 年 AI 视频生成模型横评：Elo、每秒价格与开源权重对比

The AI video leaderboard looks very different from 12 months ago. Native audio is now table stakes. Clips run 15–30 seconds. And an open-weights model sits in the top 3.

This ranking (as of Nov 13, 2026) uses one quality signal: blind human-preference Elo from the Artificial Analysis Video Arena. We pair it with API price per second and the license status of the weights. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.

Methodology

Quality: Artificial Analysis ranks models by Elo from blind pairwise votes by validated evaluators. Each vote compares 2 clips from the same prompt. We rank by the AA-Video-T2V v2.1 (with audio) board. We also report the AA-Video-I2V v2.1 score.

Anchor: Kling 3.0 1080p (Pro) is fixed at 1000 Elo on both boards. A score below 1000 means voters preferred Kling 3.0 more often.

Price: Artificial Analysis lists cost per minute on each creator’s API. It normalizes to 1080p where offered. We divide by 60 to get price per second. Vendor rates at 720p are often lower, and we flag those gaps.

Confidence: Most top-10 scores carry a ±13 to ±15 Elo interval. Models within ~25 points are statistically close. Treat ranks 3–8 as a cluster, not a strict order.

Master Comparison Table

Source: Artificial Analysis T2V and I2V leaderboards, retrieved November 13, 2026. $/sec = AA per-minute price ÷ 60. “H3 post-train” means the model was fine-tuned from MiniMax H3’s weights.

The Rankings

1. Alibaba Wan 3.0: Best overall text-to-video

• T2V Elo: 1175 (#1) · I2V Elo: 1108 (#11)
• Price: $0.20/sec at 1080p
• Weights: Closed

Wan 3.0 holds the top text-to-video spot with 7,651 votes behind it. Alibaba launched it on August 24, 2026. It generates 30-second 1080p clips from text, PDFs and slide decks, priced at $0.20 per second.

The document-to-video input is the differentiator. You can hand it a deck and get a structured video back.

The trade-off is openness. There are no downloadable weights for Wan 3.0. Wan 2.2 (July 2025) remains Alibaba’s last open video flagship. Note the I2V gap too: Wan 3.0 drops to 11th on image-to-video.

Best for: long single takes, explainer videos from existing documents.

2. Utopai Studios Utopai X: Highest-ranked post-train, no API

• T2V Elo: 1170 (#2) · I2V Elo: 1151 (#2)
• Price: No public API
• Weights: Closed

Utopai X is the only model in the top 2 on both boards. Utopai Studios, a Mountain View film studio, built it by post-training MiniMax H3. It beats the base H3 it came from.

Access runs only through Utopai’s PAI platform. PAI charges 93 credits per second, with plans starting at $15 for 1,000 credits. Developers cannot call it from a pipeline.

Best for: filmmakers working inside a studio tool. Not for programmatic use.

3. MiniMax H3: Best open-weights model

• T2V Elo: 1161 (#3) · I2V Elo: 1133 (#4)
• Price: $0.08/sec
• Weights: Open, under the MiniMax H3 Community License

H3 is the first open-weights model in the video top tier. MiniMax launched the API on July 31, 2026 and published weights on August 3. It is a 33B dense single-stream Transformer, using Qwen3-VL-32B as its encoder. It outputs video with native 32 kHz stereo audio.

Read the license before you download. The Hugging Face weights carry real limits:

• Territory: the license blocks local deployment in the US, EU, UK and South Korea.
• Resolution: only H3-Base is open, with a 768p ceiling. 2K upscaling stays API-only.
• Size: the repo holds 2 task-specific checkpoints totaling 288 GB.

Even so, H3 has become the base layer of 2026 video. Utopai X, HeyGen Video 1, fal’s H3 Max and Pika H3 Flex are all post-trains of it. AA also lists Pruna’s P-Video-2-Pro as a MiniMax collaboration.

Best for: teams outside excluded regions that want to fine-tune a frontier video model.

4. ByteDance Dreamina Seedance 2.5: Best image-to-video

• T2V Elo: 1155 (#4) · I2V Elo: 1161 (#1)
• Price: $0.569/sec on AA’s normalized basis
• Weights: Closed

Seedance 2.5 leads image-to-video. ByteDance began rolling it out on July 31, 2026 through Jimeng AI and Doubao Pro. It supports up to 30-second generations with up to 30 image, 10 video and 10 audio references.

Pricing needs care. AA lists $34.12 per minute, the highest on the board. BytePlus lists $1.156 for a 5-second 720p clip without video input, or about $0.23/sec. Confirm API access in your region before planning around it.

Best for: reference-heavy, image-led ad and character work.

5. Google Gemini Omni Flash 1.2: Best frontier price-to-quality

• T2V Elo: 1154 (#5) · I2V Elo: 1086 (#14)
• Price: $0.104/sec on AA’s basis; $0.07/sec at 720p per Google
• Weights: Closed

Gemini Omni Flash 1.2 is the newest entry, generally available since November 5, 2026. It added 3–15 second clips and a 30% price cut over version 1.1. Video output costs $12 per 1M tokens, about $0.070 per second at 720p.

It jumped 27 Elo over Omni Flash 1.1 on text-to-video. Requests to 1.1 route to 1.2 from December 3, 2026.

Best for: teams already on the Gemini API who want a top-5 model under $0.11/sec.

6. fal FLUX 3 Max: Fastest near-frontier generation

• T2V Elo: 1153 (#6) · I2V Elo: 1068 (#16)
• Price: $0.145/sec
• Weights: Closed

fal Research post-trained FLUX 3 Max with Black Forest Labs, launching it October 19, 2026. fal says a 10-second 720p clip renders in about 12 seconds. Seedance 2.5 took 277.80 seconds in fal’s own comparison.

It shares an API with fal’s H3 Max. Switching between them takes a single model-ID change.

Best for: latency-sensitive pipelines and rapid iteration.

7. HeyGen Video 1: Best value in the top 10

• T2V Elo: 1152 (#7) · I2V Elo: 1124 (#6)
• Price: $0.04/sec
• Weights: Closed

HeyGen Video 1 is a general scene generator, not an avatar engine. It is a post-train of MiniMax H3. It leads the silent (audio-off) text-to-video board at 1146 Elo.

At $0.04 per second, it costs half of base H3 for a similar score. The API model ID is heygen-video-1.

Best for: high-volume social clips and ad variants where discard rate drives cost.

8. fal MiniMax H3 Max: Fast H3 variant

• T2V Elo: 1149 (#8) · I2V Elo: 1115 (#9)
• Price: $0.16/sec
• Weights: Closed (API only)

fal Research post-trained H3 Max on the open MiniMax H3 base. It focused on prompt adherence and visual quality. fal claims it renders a 5-second clip in about 3 seconds.

It briefly led image-to-video after its August 2026 launch. Newer post-trains have since passed it.

Best for: fal users who want H3 quality without self-hosting.

9. Skywork AI SkyReels V4

• T2V Elo: 1137 (#9) · I2V Elo: 1125 (#5)
• Price: $0.35/sec
• Weights: Closed

SkyReels V4 dates from March 2026 yet still ranks top-5 on image-to-video. At $0.35 per second, it is hard to justify on text-to-video. Gemini Omni Flash 1.2 scores higher for less than a third of the price.

10. Black Forest Labs FLUX 3

• T2V Elo: 1137 (#10) · I2V Elo: 1071 (#15)
• Price: $0.29/sec
• Weights: Closed today; open Dev version promised

FLUX 3 is BFL’s first video model, with 20-second clips and native audio. One backbone also handles robot action prediction. BFL says it will release open-weight versions later this year. Until FLUX 3 Dev ships, treat it as API-only.

Open-Weights Leaderboard

MiniMax H3 family: Pika H3 Flex shipped October 29, 2026. It cuts H3 to 6 sampling steps. Local inference still needs an H100 80GB and about 215 GB of downloads. Video Rebirth’s HyperFlow H3 Flash compresses H3 to 13.7B. It runs from a 9.1 GB GGUF on a 24 GB GPU. Its sibling, Hyperflow H3, ranks #7 on image-to-video at 1120 Elo.

LTX-2.5: Lightricks released the 22B model on August 11, 2026. It adds native multishot generation and joint audio. Under the license, entities with $10M+ annual revenue need a paid license. It trails the H3 family by roughly 200 Elo. But it gates by revenue, not territory. For US and EU teams, that makes it a practical default. See our LTX-2.5 launch coverage.

Apache 2.0 options: Wan 2.2 and HunyuanVideo 1.5 remain the go-to picks for permissive licensing or a single consumer GPU. Neither appears on the current AA boards.

Price per Second: Where the Value Is

Quality and price no longer track each other. The 3 cheapest top-tier models are all H3 derivatives or H3 itself.

Value = (Elo − 1000) ÷ price per second. Higher is better. Illustrative only; it ignores resolution and clip-length differences.

At the top, Wan 3.0 buys you about 23 Elo over HeyGen Video 1. It costs 5x more per second. Whether that matters depends on how many takes you discard.

What About Veo, Kling 4.0, Sora and Runway?

• Google Veo 3.1 scores 964 on text-to-video. That is below the Kling 3.0 anchor. Veo 3.1 Lite matches base H3 on price at $0.08/sec but trails it by 213 Elo.
• Kling 4.0 became generally available on klingai.com with 30-second native clips. Kuaishou had not announced API availability or pricing at last check. It does not yet appear on the AA boards.
• OpenAI Sora 2 and Runway Gen-4.5 are not on the current AA with-audio boards. We excluded them rather than mix in non-comparable scores.

How to Choose

1. Highest text-to-video quality, API access: Wan 3.0.
2. Image-to-video and heavy references: Seedance 2.5, budget permitting.
3. Best quality per dollar: HeyGen Video 1 or Gemini Omni Flash 1.2.
4. Speed for iteration: FLUX 3 Max or H3 Max on fal.
5. Self-hosting outside the US/EU/UK/South Korea: MiniMax H3 or Pika H3 Flex.
6. Self-hosting in the US or EU: LTX-2.5, or Wan 2.2 for Apache 2.0.

Key Takeaways

• Alibaba’s Wan 3.0 leads Artificial Analysis text-to-video with audio at 1175 Elo, at $0.20/sec.
• ByteDance’s Seedance 2.5 leads image-to-video at 1161 Elo, but costs the most on the board.
• MiniMax H3 is the top open-weights model at #3, and at least 4 top-14 models are post-trains of it.
• HeyGen Video 1 and Pika H3 Flex deliver near-frontier Elo for $0.035–$0.04 per second.
• Google’s Veo 3.1 has slipped below the Kling 3.0 anchor on both blind-vote boards.
• You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.

Sources

• Artificial Analysis — Text-to-Video Leaderboard (AA-Video-T2V v2.1)
• Artificial Analysis — Image-to-Video Leaderboard (AA-Video-I2V v2.1)
• DataNorth — Alibaba launches Wan3.0
• The Next Web — Alibaba launches Wan3.0
• AlphaSignal — Utopai X beats the MiniMax H3 model it was built on
• Hugging Face — MiniMaxAI/MiniMax-H3
• mer.vin — MiniMax H3 open weights blocked in US/EU/UK
• AI Weekly — MiniMax posts H3 to Hugging Face
• RuntimeWire — MiniMax H3 open weights and license
• Leadde — Seedance 2.5 pricing
• CometAPI — Seedance 2.5 API pricing
• TokenCost — Gemini Omni 1.2 Flash pricing
• Tech Insider — Gemini Omni 1.2 Flash price cut
• fal — Introducing FLUX 3 Max and FLUX 3 Max page
• OrcaRouter — HeyGen Video 1 launch
• Kompozy — HeyGen Video tops silent T2V board
• fal — H3 Max and launch post
• VentureBeat — Black Forest Labs launches FLUX 3
• DataNorth — Black Forest Labs releases FLUX 3
• Hugging Face — Pika-Labs/Pika-H3-Flex
• AlphaSignal — Pika drops H3 Flex
• ComfyUI Wiki — HyperFlow H3 Flash
• LTX — Introducing LTX-2.5
• RITS — LTX-2.5 license and specs
• Let’s Data Science — Kling 4.0 opens to all users
• No Code MBA — Kling 4.0 release tracker
• Pinggy — Best video generation models 2026

The post Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights appeared first on MarkTechPost.