Utopai X: Film Studio Fine-Tunes Chinese Model into Top 2 Video AI
Utopai Studios fine-tuned MiniMax H3 to take 2nd place on Artificial Analysis text-to-video leaderboard. Here is what it costs and why it matters.
👁 113 reads
The world’s second-best AI video model according to blind comparisons on Artificial Analysis does not come from Google, OpenAI, or a Chinese tech giant. It was created by Utopai Studios - a Mountain View film studio that fine-tuned MiniMax’s open-weights model and ended up outperforming the original. Today, Utopai X holds an Elo rating of 1149, sitting just 7 points behind Alibaba’s Wan 3.0. I have gathered all the figures, pricing, and hands-on details in my Utopai X guide.
Utopai X is the highest-ranked model from a U.S.-based company on Artificial Analysis’ Text-to-Video Leaderboard with Audio.
- Utopai Studios (@UtopaiStudios) September 30, 2026
What happened
On September 30, 2026, Utopai Studios unveiled a new version of its PAI platform alongside its first proprietary video model, Utopai X. On the same day, Artificial Analysis, which runs the most widely cited model leaderboard in the industry, ranked it in 2nd place on its text-to-video with audio leaderboard. In its model card, AA added one key note: “based on MiniMax H3.”
The ranking is based on blind comparisons. Evaluators view two clips generated from the same prompt without model names and pick the better output. By October 11, Utopai X had accumulated 6,071 comparisons. Here was the top of the leaderboard on that date:
| Rank | Model | Developer | Elo |
|---|---|---|---|
| 1 | Wan 3.0 | Alibaba (China) | 1156 |
| 2 | Utopai X | Utopai Studios (US) | 1149 |
| 3 | Seedance 2.5 | ByteDance (China) | 1143 |
| 4 | MiniMax H3 | MiniMax (China) | 1137 |
| 5 | MiniMax H3 Max | MiniMax and fal.ai | 1130 |
| 6 | FLUX 3 | Black Forest Labs (Germany) | 1128 |
Utopai emphasizes that it is the highest-rated model from a US company. That is true, but with a twist: the core engine originated in Shanghai.
Chinese engine, American fine-tuning
MiniMax launched H3 in late July 2026 and released the base model with open weights. Utopai Studios took that foundation and fine-tuned it specifically for on-set production needs: physics (mass and object contact), camera motion, material lighting, and action-synchronized audio. The gains are measurable - Utopai X scores 12 points higher than base H3 and also surpasses H3 Max, a version built on the same foundation by fal.ai.
In its launch-day category breakdown, Artificial Analysis awarded Utopai X 1st place in physics and audio synchronization, and 2nd place in lighting and camera movement, behind Wan 3.0. The studio showcased this across six arena comparisons: a boat being pushed into water, a plush toy under lamplight, a sunlit construction site, a robot inside a cockpit, two presenters, and a rapid pan shot. I have documented all of them in What Utopai X Can Do.
Fine-tuning an open-weight base and keeping the resulting weights proprietary did raise eyebrows. On X, community members questioned whether the studio built upon open-source code only to close it off. Utopai X offers neither open weights nor an API.
Who is Utopai Studios
This is not your average model research startup. Founded in 2022 as Cybever by Cecilia Shen and Jie Yang, the company rebranded in August 2025. It produces original films and series, including the historical epic “Cortés” and the sci-fi series “Space Nation.” In May 2026, an investment from Carmelo Anthony’s fund valued the company at $1 billion. The PAI platform hosting Utopai X launched in March 2026 and, according to the company, reached $11 million in annual recurring revenue within 60 days.
Pricing and access
- Available exclusively in PAI. The model is not offered via an API, nor is it listed on OpenRouter or reAPI - confirmed via access checks on October 11.
- 93 credits per second of 1080p video. The $15/month Basic plan provides 1,000 credits, translating to roughly 10 seconds of Utopai X video.
- $60 to $84 per minute depending on the subscription tier, assuming every clip succeeds on the first attempt. For context: Wan 3.0 costs $12 per minute on the AA leaderboard, while MiniMax H3 costs $4.80.
- Global checkout. The site and pricing page are functional, with all tiers billed in USD. Utopai does not publish an official list of supported countries, so you will want to verify availability for your region directly.
- Beware of impostors. The domain utopaix.video is unaffiliated with the studio and does not provide access to the genuine model.
Why this matters
Until now, video AI leaderboards have been dominated by organizations training foundation models from scratch: Alibaba, ByteDance, Google, MiniMax, and Kuaishou. Utopai X demonstrates a viable alternative - a studio with deep filmmaking expertise can take an open-weight base, adapt it to real-world production requirements, and compete at the very top. That is a promising signal for independent studios and a strong proof point for open weights, which made Utopai’s achievement possible.
For independent creators and developers, the practical takeaway is straightforward. If your workflow relies on API calls or automated pipelines, models like Wan 3.0, H3 Max, and FLUX 3 remain far more cost-effective and readily accessible. Utopai X makes the most sense for production teams already running end-to-end projects inside PAI. You can explore the detailed head-to-head metrics in Utopai X vs Wan 3.0 vs MiniMax H3, and check my tests of alternative models on matching prompts in the Video Arena.
More on this from QUANTSLANT
magazine·Gemini 4 Argon Price: What Google’s New Model Costs in 2026Gemini 4 Argon starts at $2 per million input tokens and $10 per million output tokens during the introductory period.

