HappyHorse 1.1: Alibaba Video Generator Review, Pricing, Rank

👁 117 reads

Continuously updated. Current as of October 11, 2026.

HappyHorse 1.1 is Alibaba’s AI video generation model from its ATH unit, launched on June 22, 2026. On the Artificial Analysis leaderboard, it ranks 12th in text-to-video (1042 Elo) and 8th in image-to-video (1104 Elo), sitting noticeably behind Alibaba’s own Wan 3.0. It produces 3-15 second clips at up to 1080p resolution with native audio. I tested it across 7 test scenes via OpenRouter: $0.494 for a 5-second 720p clip, solid camera movement, and consistent characters, but weak non-English text and Polish speech rendering.

What is HappyHorse and who makes it?

HappyHorse is a family of AI video models with native audio developed by Alibaba alongside its well-known Wan line. On the Artificial Analysis leaderboard, the creator is listed as “Alibaba-ATH”. ATH stands for Alibaba Token Hub, a new business unit established on March 16, 2026 (according to Caixin). According to the Picsart tech blog, the model was built by the Future Life Lab team inside the Taotian Group (the business group behind Taobao and Tmall). Website assets on happyhorse.com are also served from a “future-life-web” directory hosted on Alibaba’s servers.

The anonymous debut of version 1.0

HappyHorse started as an industry puzzle. Here is how events unfolded:

• April 7, 2026: Artificial Analysis introduced an anonymous model to its video arena, which immediately claimed the #1 spot.
• April 10, 2026: AA revealed the mysterious model was developed by Alibaba. Alibaba’s Hong Kong-listed shares gained 2.12% that day (Caixin, CNBC).
• June 22, 2026: Release of HappyHorse 1.1, officially announced by Alibaba Group and Alibaba Cloud with a 40% launch discount for two weeks. VentureBeat reported it reached #2 globally at launch.
• June 23, 2026: Launch of the HorsePower AI Video Competition alongside Huajing Entertainment Group, offering prizes up to 1 million RMB.
• July 6, 2026: End of the 40% launch promotion.

Why did it slip from #2 in June to #12 today? Since then, Wan 3.0, Seedance 2.5, MiniMax H3, and FLUX 3 have launched. In addition, AA overhauled its text-to-video benchmark methodology on September 30 (v2.0, testing at 1080p with audio). Elo ratings from June and October operate on distinct scales and cannot be compared directly.

What changed in version 1.1?

According to the Alibaba Cloud blog from June 23, version 1.1 delivers improvements in:

• multi-reference comprehension and visual consistency for characters and products in reference mode,
• motion fluidity in dynamic action scenes,
• prompt adherence, visual detail, and audiovisual synchronization.

In terms of concrete technical specs: 480p resolution was introduced, and 1080p generation dropped from $0.24 to $0.18 per second in the Singapore region (a 25% price cut). Editing existing video remained exclusive to version 1.0 (happyhorse-1.0-video-edit).

Artificial Analysis leaderboard rankings

The Artificial Analysis leaderboard uses blind pairwise comparisons: human raters evaluate two anonymously rendered clips generated from the exact same prompt and vote for the superior output. I pulled both tables on October 11, 2026.

In text-to-video, HappyHorse 1.1 is in a statistical tie with nearby contenders: the confidence intervals for Agnes Video 2.5 (10th), Grok Imagine 1.5 (11th), and Wan 2.7 (13th) overlap with its score. It sits more than 100 Elo points behind the category leaders. In image-to-video, the picture is clearer: 8th place, significantly ahead of version 1.0, though trailing Wan 3.0 in 7th place by 60 points.

The chart baseline starts at 1000 Elo (with 1200 representing a full bar) to highlight relative differences.

Why it scores higher in Image-to-Video

Directly comparing Elo points across the two tables is invalid because each uses a different pool of models and arena matchups. However, the ordinal ranks (8th vs 12th) are informative. My observation: in image-to-video, scene composition, lighting, and art style are predetermined by the starting frame; the model is judged primarily on realistic motion. Dynamic motion happens to be HappyHorse’s strong suit, which was evident in my own testing. In text-to-video, the generator must synthesize the entire scene from scratch, and under AA’s v2.0 methodology (judged at 1080p with synchronized audio), flaws across composition, physics, and sound are exposed simultaneously.

Alibaba’s Wan 3.0 leads HappyHorse

Alibaba has two competing models developed by distinct internal teams. Wan 3.0 (released August 19, 2026) leads the text-to-video leaderboard at 1156 Elo (114 points ahead of HappyHorse 1.1) and holds 7th place in image-to-video at 1164 Elo. On AA’s pricing index, Wan 3.0 runs $12.00 per minute. If you are deciding between them purely on generation quality, the data clearly favors Wan.

HappyHorse 1.1 technical specifications

Specifications sourced from Alibaba Cloud Model Studio documentation (checked October 11, 2026). When writing prompts with reference images, you cite images as “[Image 1]”, on happyhorse.com as “@Image1”, and on fal.ai as character1 through character9.

HappyHorse 1.1 pricing and API costs

Rates verified on October 11, 2026. Costs for a 5-second clip and 1 minute represent calculations from the per-second rate.

The most economical API routes are OpenRouter at 720p and the Frankfurt region in Alibaba Cloud. Alibaba Cloud includes a free tier of 10 seconds per model valid for 90 days. The Artificial Analysis leaderboard benchmarks HappyHorse 1.1 at $10.80 per minute for text-to-video and $9.90 for image-to-video.

happyhorse.com web subscription plans

Based on the Pro tier’s stated capacity, cost per 5-second 720p video averages roughly $0.27, which is cheaper than direct API calls. Annual plans were not audited. Source: happyhorse.com/subscription (October 11, 2026).

Global availability and access

• happyhorse.com: Accessible worldwide with billing in USD. Note that terms of service documentation was inaccessible (/terms returned a 404 error at testing), so confirm commercial terms inside your account before commercial deployment.
• Alibaba Cloud Model Studio: Offers an international deployment in the Frankfurt region. Ensure your selected model, service endpoint, and API key belong to the identical geographic region, or requests will fail.
• OpenRouter, fal.ai, Replicate, and WaveSpeedAI: Fully operational without an Alibaba Cloud account. Together AI supports the text-to-video endpoint.

Beware of counterfeit domains. Multiple third-party sites spoof the model’s branding (such as happyhorse-ai.ai, happyhorse.app, or tryhappyhorse.com) with separate subscription checkouts. The official site is happyhorse.com (alongside its .cn portal), and the official X profile is @HappyHorseATH.

Open weights: none available

Neither HappyHorse 1.1 nor version 1.0 offer open weights. Portals promoting “HappyHorse open source Apache 2.0” (e.g., happyhorse.me, happyhorse-model.com) and associated GitHub repositories are unofficial copycats. Avoid downloading files from them. You cannot run this model on local hardware. If you need an Alibaba open-weights video model for local GPU execution, use Wan 2.2 as outlined in our guide to Wan AI and our directory of local AI models.

Hands-on benchmark test

On October 11, 2026, I generated 7 benchmark clips via OpenRouter (alibaba/happyhorse-1.1, 720p, 5 seconds, seed 20261008). Each clip cost $0.494 and required 96 to 112 seconds to generate. Scenes 1 through 5 replicated prompts from my standardized Video Arena benchmark from October 8, assessed using the same scoring protocol (four extracted frames plus full video playback, rated 0-10).

The average score across the four standard non-speech Arena scenes is 6.1. That matches ByteDance’s Seedance 2.5 (6.0), while trailing Wan 3.0 (9.1) and MiniMax H3 Max (8.25) by a wide margin. Reference-to-video mode was not tested due to API balance constraints.

For prompts and frame-by-frame breakdowns of all 7 test scenes, read our companion breakdown: What HappyHorse 1.1 can do. For deep-dive audio analysis, see our speech and text rendering test.

How to get started step by step

Using the web interface

1. Visit happyhorse.com (confirm the exact URL without extra subdomains) and register an account.
2. Select “HH 1.1” and your preferred generation mode: Text, “First frame” (image-to-video), or “Reference to video” (up to 9 reference stills).
3. Enter your prompt in English. Place dialogue lines inside quotation marks.
4. Select your duration (3-15 seconds), aspect ratio, and resolution. Free tiers apply a visible watermark.
5. Download your rendered clip from the asset library.

Using the OpenRouter API

1. Create an account at OpenRouter and add balance. A 5-second 720p generation costs $0.494.
2. Submit a POST https://openrouter.ai/api/v1/videos request with your bearer token header: Authorization: Bearer YOUR_KEY.
3. Poll the returned polling_url every few seconds until the job state reaches completed.
4. Retrieve your video URL from unsigned_urls[0]. The exact cost appears in the usage.cost response field.

Here is the exact request payload used in my image-to-video benchmark (API key omitted, image string truncated):

{
  "model": "alibaba/happyhorse-1.1",
  "prompt": "Slow camera push-in on the green ceramic mug. Hot black coffee is poured into the mug from above in a thin stream...",
  "duration": 5,
  "resolution": "720p",
  "aspect_ratio": "16:9",
  "seed": 20261008,
  "frame_images": [
    {"type": "image_url",
     "image_url": {"url": "data:image/jpeg;base64,..."},
     "frame_type": "first_frame"}
  ]
}

For text-to-video requests, simply omit the frame_images parameter. Instead of base64 strings, you can provide a direct publicly accessible image URL. On Alibaba Cloud, model slugs are happyhorse-1.1-t2v, -i2v, and -r2v; verify that your API credentials originate from the Frankfurt region. For wider ecosystem rate comparisons, review our overview of how much AI video generation costs per second.

Comparison with rival video models

Benchmark data sourced from Artificial Analysis, verified October 11, 2026. MiniMax H3 Max places higher across both leaderboards while costing less than half as much per minute of output. HappyHorse 1.1 is in a statistical tie with Grok Imagine 1.5 in both rankings, though HappyHorse is cheaper in text-to-video mode. For head-to-head score breakdowns on identical scenes, see our article on HappyHorse 1.1 vs Wan 3.0 vs Seedance vs MiniMax, or review our broad market guide to AI video generators.

Who should use HappyHorse 1.1?

• Creators focusing on image-to-video with intentional camera movement. Orbit and push-in tracking shots stayed smooth without warping the source object.
• Multi-shot continuity with a single character. Following a hard cut, the model preserved facial structure, glasses, and outfit details cleanly.
• Projects requiring English dialogue generation. The spoken English line tested here achieved a zero word error rate under Whisper transcription.

Who should avoid it?

• Non-English voice generation. When we tested it in Polish, the model mispronounced half the words.
• Precise in-frame typography. The model consistently drops characters and misinterprets accent marks.
• Teams seeking maximum visual fidelity per dollar. Wan 3.0 and MiniMax H3 Max deliver significantly better outputs in benchmarks and practical scoring.
• Local pipeline workflows. Model weights are strictly closed source.

Frequently asked questions

Is HappyHorse 1.1 free to use?

happyhorse.com offers a free plan featuring daily recurring credits and 2 concurrent generation slots. Alibaba Cloud includes 10 free seconds per model valid for 90 days. Unmetered, permanent free API access does not exist.

How much does a 5-second video clip cost?

On OpenRouter at 720p resolution, a 5-second clip cost me $0.494. Alibaba Cloud (Frankfurt region) bills approximately $0.62 for 720p, while fal.ai and Replicate charge $0.70. Under happyhorse.com’s Pro subscription, the effective cost drops to roughly $0.27 per clip.

Does HappyHorse support foreign language speech like Polish?

Performance is poor. When we tested it in Polish using a classic tongue-twister, 6 out of 12 words contained pronunciation errors. Alibaba has not published an official language support matrix, and third-party hosting partners do not list non-English dialogue capabilities.

Are HappyHorse model weights open source?

No. Sites offering “HappyHorse open source” files are unauthorized copycats. Alibaba’s open-weights alternative is the Wan 2.2 series.

How does HappyHorse compare to Wan?

They are independent architectures developed by different teams inside Alibaba. Wan 3.0 holds 1st place in text-to-video and 7th in image-to-video, whereas HappyHorse 1.1 ranks 12th and 8th. In our benchmark tests, Wan 3.0 averaged 9.1/10 compared to 6.1/10 for HappyHorse 1.1.

Can HappyHorse 1.1 be accessed globally?

Yes. The happyhorse.com portal is open internationally, and Alibaba Cloud hosts endpoints in Frankfurt, Beijing, Singapore, and Virginia. The simplest integration path is via API routers like OpenRouter, fal.ai, or Replicate.

Can you edit existing video clips in version 1.1?

No. Video-to-video editing workflows are only available in the legacy happyhorse-1.0-video-edit endpoint.