2026年AI视频生成模型榜:Elo、每秒价格、开放权重
AI 视频排行榜与 12 个月前相比已大不相同。原生音频如今已是入场门槛。片段时长为 15–30 秒。而且有一款开放权重模型跻身前三。
本排名(截至 2026 年 11 月 13 日)采用一项质量信号:来自 Artificial Analysis Video Arena 的盲测人类偏好 Elo。我们将其与 API 每秒价格以及权重的许可状态并列呈现。你可以通过我们的 GPULeaderboard.com Chrome 扩展在这里查看部署选项与规格。
质量:Artificial Analysis 依据经过验证的评估者的盲测成对投票所得 Elo 对模型进行排名。每次投票比较同一提示词生成的两个片段。我们以 AA-Video-T2V v2.1(含音频)榜单进行排名,同时也报告 AA-Video-I2V v2.1 分数。
基准锚点:Kling 3.0 1080p(Pro)在两个榜单上均固定为 1000 Elo。分数低于 1000 意味着投票者更常偏好 Kling 3.0。
价格:Artificial Analysis 列出各家创作者 API 的每分钟成本,并在提供 1080p 时统一折算为 1080p。我们除以 60 得到每秒价格。厂商的 720p 价格通常更低,我们会标注这些差异。
置信度:前十名中的多数分数带有 ±13 至 ±15 的 Elo 区间。差距在约 25 分以内的模型在统计上不相上下。第 3 至第 8 名应视为一个梯队,而非严格的排序。
来源:Artificial Analysis T2V 与 I2V 排行榜,检索于 2026 年 11 月 13 日。美元/秒 = AA 每分钟价格 ÷ 60。“H3 post-train”指该模型基于 MiniMax H3 的权重微调而来。
Wan 3.0 以 7,651 票位居文生视频榜首。阿里巴巴于 2026 年 8 月 24 日发布该模型。它可根据文本、PDF 和幻灯片生成 30 秒的 1080p 片段,价格为每秒 0.20 美元。
文档转视频输入是它的差异化所在。你可以交给它一份演示文稿,然后拿回一段结构化的视频。
代价是开放性。Wan 3.0 没有可下载的权重。Wan 2.2(2025 年 7 月)仍是阿里巴巴最后一款开放权重的视频旗舰模型。还要注意图生视频(I2V)上的差距:Wan 3.0 在图像生成视频榜单上跌至第 11 位。
最适合:长单镜头生成、基于已有文档的解说视频。
Utopai X 是唯一在两个榜单上都跻身前 2 的模型。位于山景城的影视工作室 Utopai Studios 通过对 MiniMax H3 进行后训练打造了它。它的表现超过了它所基于的基础版 H3。
只能通过 Utopai 的 PAI 平台访问。PAI 每秒收费 93 积分,套餐起步价为 15 美元 1,000 积分。开发者无法从流水线中调用它。
最适合:在工作室工具内工作的影视创作者。不适合程序化调用。
H3 是视频领域第一梯队中首个开放权重模型。MiniMax 于 2026 年 7 月 31 日上线 API,并于 8 月 3 日公开权重。它是一个 33B 稠密单流 Transformer,使用 Qwen3-VL-32B 作为编码器,可输出带原生 32 kHz 立体声音频的视频。
下载前请先阅读许可协议。Hugging Face 上的权重带有实实在在的限制:
即便如此,H3 已成为 2026 年视频领域的基础层。Utopai X、HeyGen Video 1、fal 的 H3 Max 和 Pika H3 Flex 都是基于它的后训练版本。AA 还将 Pruna 的 P-Video-2-Pro 列为与 MiniMax 合作的成果。
最适合:位于排除区域之外、希望微调前沿视频模型的团队。
Seedance 2.5 在图像生成视频方面领先。字节跳动于 2026 年 7 月 31 日通过 Jimeng AI 和 Doubao Pro 开始推送该模型。它支持最长 30 秒的生成,最多可引用 30 张图像、10 段视频和 10 段音频。
定价需谨慎。AA 列出的是每分钟 34.12 美元,为榜单最高。BytePlus 列出的是无视频输入时一段 5 秒 720p 视频 1.156 美元,约合每秒 0.23 美元。在据此做规划前,请先确认你所在区域能否使用该 API。
最适合:大量依赖参考素材、以图像为主导的广告和角色类工作。
Gemini Omni Flash 1.2 是最新入榜者,自 2026 年 11 月 5 日起全面可用。它新增了 3–15 秒片段能力,并较 1.1 版本降价 30%。视频输出价格为每 100 万 Token 12 美元,720p 下约合每秒 0.070 美元。
在文生视频方面,它较 Omni Flash 1.1 提升了 27 个 Elo。自 2026 年 12 月 3 日起,发往 1.1 的请求将路由至 1.2。
最适合:已经在使用 Gemini API、希望以低于每秒 0.11 美元的价格用上前五名模型的团队。
fal Research 与 Black Forest Labs 合作对 FLUX 3 Max 进行了后训练,于 2026 年 10 月 19 日发布。fal 表示,一段 10 秒的 720p 视频约 12 秒即可渲染完成。在 fal 自己的对比中,Seedance 2.5 耗时 277.80 秒。
它与 fal 的 H3 Max 共用同一套 API,两者之间切换只需改一个模型 ID。
适用场景:对延迟敏感的流水线以及快速迭代。
HeyGen Video 1 是通用场景生成器,而非数字人引擎。它是对 MiniMax H3 的后训练版本。它以 1146 Elo 领跑无声(关闭音频)文生视频榜单。
价格是每秒 0.04 美元,在得分相近的情况下只有基础版 H3 的一半。API 模型 ID 为 heygen-video-1。
适用场景:大批量社交短视频和广告变体,此类场景下废片率直接决定成本。
fal Research 在开放的 MiniMax H3 基础模型上后训练出 H3 Max,重点优化提示词遵循度和画面质量。fal 声称它渲染一段 5 秒视频约需 3 秒。
它在 2026 年 8 月发布后曾一度领跑图生视频,但此后已被更新的后训练版本超越。
最适合:想要 H3 质量但不想自托管的 fal 用户。
SkyReels V4 发布于 2026 年 3 月,但至今仍在图生视频榜上位列前五。以每秒 0.35 美元的价格,它在文生视频上很难说得过去。Gemini Omni Flash 1.2 得分更高,价格却不到它的三分之一。
FLUX 3 是 BFL 的首个视频模型,支持 20 秒片段和原生音频。同一个主干网络还能处理机器人动作预测。BFL 表示将于今年晚些时候发布开放权重版本。在 FLUX 3 Dev 发布之前,只能把它当作仅提供 API 的模型。
MiniMax H3 家族:Pika H3 Flex 于 2026 年 10 月 29 日发布。它将 H3 缩减到 6 个采样步。本地推理仍需一块 H100 80GB,下载量约 215 GB。Video Rebirth 的 HyperFlow H3 Flash 将 H3 压缩到 13.7B。它可在 24 GB GPU 上以 9.1 GB 的 GGUF 运行。其同门 Hyperflow H3 以 1120 Elo 在图生视频榜上排名第 7。
LTX-2.5:Lightricks 于 2026 年 8 月 11 日发布这款 22B 模型。它新增了原生多镜头生成和联合音频。根据其许可证,年营收 1000 万美元以上的实体需要购买付费许可。它比 H3 家族落后约 200 Elo。但它的限制条件是营收,而非地域。对美国与欧洲团队来说,这使它成为实用的默认选择。参见我们的 LTX-2.5 发布报道。
Apache 2.0 选项:Wan 2.2 和 HunyuanVideo 1.5 仍是宽松许可或单张消费级 GPU 场景的首选。两者都未出现在当前的 AA 榜单上。
质量与价格不再同步。最便宜的 3 个顶级模型全都是 H3 的衍生版本或 H3 本身。
性价比 =(Elo − 1000)÷ 每秒价格。数值越高越好。仅供参考;它忽略了分辨率和片段长度的差异。
排在榜首的 Wan 3.0 相比 HeyGen Video 1 大约高出 23 个 Elo,但每秒成本高出 5 倍。这是否重要,取决于你要丢弃多少条生成结果。
《2026 年最佳 AI 视频生成模型:Elo、每秒价格、开放权重》一文最初发表于 MarkTechPost。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights)
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
| Rank | Model | Developer | T2V Elo (audio) | I2V Elo (audio) | $/sec | 10-sec clip | Open weights |
|---|---|---|---|---|---|---|---|
| 1 | Wan 3.0 | Alibaba | 1175 | 1108 | $0.200 | $2.00 | No |
| 2 | Utopai X | Utopai Studios (H3 post-train) | 1170 | 1151 | No API | — | No |
| 3 | MiniMax H3 (768p) | MiniMax | 1161 | 1133 | $0.080 | $0.80 | Yes (restricted) |
| 4 | Dreamina Seedance 2.5 | ByteDance Seed | 1155 | 1161 | $0.569 | $5.69 | No |
| 5 | Gemini Omni Flash 1.2 | 1154 | 1086 | $0.104 | $1.04 | No | |
| 6 | FLUX 3 Max | fal × Black Forest Labs | 1153 | 1068 | $0.145 | $1.45 | No |
| 7 | HeyGen Video 1 | HeyGen (H3 post-train) | 1152 | 1124 | $0.040 | $0.40 | No |
| 8 | MiniMax H3 Max | fal (H3 post-train) | 1149 | 1115 | $0.160 | $1.60 | No |
| 9 | SkyReels V4 | Skywork AI | 1137 | 1125 | $0.350 | $3.50 | No |
| 10 | FLUX 3 | Black Forest Labs | 1137 | 1071 | $0.290 | $2.90 | No (Dev promised) |
| — | Seedance 2.0 | ByteDance Seed | 1136 | 1140 | $0.374 | $3.74 | No |
| — | P-Video-2-Pro | Pruna AI × MiniMax | 1130 | 1116 | $0.095 | $0.95 | No |
| — | Pika H3 Flex | Pika (H3 post-train) | 1124 | 1114 | $0.035 | $0.35 | Yes (restricted) |
| — | Kling 3.0 1080p (Pro) | Kuaishou | 1000 (anchor) | 1000 (anchor) | $0.168 | $1.68 | No |
| — | Veo 3.1 | 964 | 955 | $0.400 | $4.00 | No | |
| — | Veo 3.1 Lite | 948 | 931 | $0.080 | $0.80 | No | |
| — | LTX-2.5 Fast | Lightricks | 943 | 911 | $0.130 | $1.30 | Yes |
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
| Model | T2V Elo | License | Key limits |
|---|---|---|---|
| MiniMax H3 (768p) | 1161 | MiniMax H3 Community License | No deployment in US/EU/UK/South Korea; 768p ceiling |
| Pika H3 Flex | 1124 | MiniMax H3 Community License | LoRA adapters on H3; same territory terms |
| LTX-2.5 Fast | 943 | LTX-2.x Community License | Paid license at $10M+ annual revenue |
| LTX-2.5 Pro | 938 | LTX-2.x Community License | Same as above |
| Wan 2.2 | Not ranked | Apache 2.0 | Older (2025); no native audio in base release |
| HunyuanVideo 1.5 | Not ranked | Tencent Hunyuan Community License | 8.3B; aimed at single consumer GPUs |
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.
| Model | T2V Elo | $/sec | Elo above anchor per $1/sec |
|---|---|---|---|
| HeyGen Video 1 | 1152 | $0.040 | 3,800 |
| Pika H3 Flex | 1124 | $0.035 | 3,543 |
| MiniMax H3 (768p) | 1161 | $0.080 | 2,012 |
| Gemini Omni Flash 1.2 | 1154 | $0.104 | 1,476 |
| P-Video-2-Pro | 1130 | $0.095 | 1,368 |
| FLUX 3 Max | 1153 | $0.145 | 1,055 |
| Wan 3.0 | 1175 | $0.200 | 875 |
| Seedance 2.5 | 1155 | $0.569 | 273 |
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
- Highest text-to-video quality, API access: Wan 3.0.
- Image-to-video and heavy references: Seedance 2.5, budget permitting.
- Best quality per dollar: HeyGen Video 1 or Gemini Omni Flash 1.2.
- Speed for iteration: FLUX 3 Max or H3 Max on fal.
- Self-hosting outside the US/EU/UK/South Korea: MiniMax H3 or Pika H3 Flex.
- 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.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。