Source虎嗅•Stalecollected in 3h
AI Video Hits a Two-Yuan-Per-Second Price Wall
#video-gen#inference-costs#model-comparisonminimax-h3-and-seedance-2.5minimax h3seedance 2.5bytedance
The models look impressive, but RMB 2 per second may redefine AI-video unit economics.
30-Second TL;DR
What Changed
MiniMax H3 and Seedance 2.5 were released within a short period and are being compared for their video-generation quality.
Why It Matters
High inference costs could limit experimentation, user acquisition, and production volume for AI-video startups and creators. Providers that reduce per-second costs or offer efficient batch and subscription pricing may gain a stronger position.
What To Do Next
Benchmark MiniMax H3 and Seedance 2.5 on your own clips, measuring cost per usable second and rejection rate before choosing a production API.
Who should care:Creators & Designers
Key Points
- •MiniMax H3 and Seedance 2.5 were released within a short period and are being compared for their video-generation quality.
- •The reported generation cost is approximately RMB 2 per second, creating a significant barrier for casual users and high-volume creators.
- •The debate highlights the tension between rapid improvements in AI-video quality and the economics of large-scale generation.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •The RMB 2 per second pricing model is largely driven by the massive inference costs associated with high-parameter diffusion transformer (DiT) architectures required for consistent 1080p video generation.
- •Industry analysts note that current AI video generation requires significant GPU cluster utilization, specifically H100/H800 equivalents, which remain in short supply and high demand in the Chinese market.
- •Beyond raw generation costs, companies are facing 'token-to-video' latency challenges, where the time-to-first-frame (TTFF) remains a bottleneck for real-time interactive applications.
- •Major cloud providers in China are beginning to offer subsidized API credits for developers to offset these high per-second costs, aiming to capture market share in the nascent AI video ecosystem.
- •The 'two-yuan wall' has accelerated the development of model distillation techniques, where companies are attempting to create 'lite' versions of H3 and Seedance 2.5 to reduce compute overhead by 30-50%.
Competitor Analysis
MiniMax H3
- Pricing Model
- ~RMB 2/sec
- Key Strength
- High temporal consistency
- Target Audience
- Professional Creators
Seedance 2.5
- Pricing Model
- ~RMB 2/sec
- Key Strength
- Motion realism/physics
- Target Audience
- Enterprise/Marketing
Kling AI
- Pricing Model
- Variable/Credit
- Key Strength
- Long-form generation
- Target Audience
- General Public
Sora (OpenAI)
- Pricing Model
- N/A (Private)
- Key Strength
- World simulation
- Target Audience
- Research/Enterprise
| Model | Pricing Model | Key Strength | Target Audience |
|---|---|---|---|
| MiniMax H3 | ~RMB 2/sec | High temporal consistency | Professional Creators |
| Seedance 2.5 | ~RMB 2/sec | Motion realism/physics | Enterprise/Marketing |
| Kling AI | Variable/Credit | Long-form generation | General Public |
| Sora (OpenAI) | N/A (Private) | World simulation | Research/Enterprise |
Technical Deep Dive
- Both MiniMax H3 and Seedance 2.5 utilize advanced Diffusion Transformer (DiT) architectures, moving away from traditional U-Net structures to improve global coherence.
- The models employ latent space compression techniques to reduce the dimensionality of video data before processing, which is critical for managing memory bandwidth.
- Implementation involves multi-stage generation pipelines where a base model generates low-resolution keyframes, followed by temporal upsampling and spatial super-resolution modules.
- Inference optimization relies heavily on custom CUDA kernels and FP8 quantization to maximize throughput on NVIDIA hardware.
Future ImplicationsAI analysis grounded in cited sources
AI video generation costs will drop below RMB 0.5 per second by Q2 2027.
The maturation of specialized AI inference chips and the widespread adoption of model distillation will significantly reduce the compute-per-pixel requirement.
Enterprise adoption will shift toward hybrid local-cloud deployment models.
To bypass the high per-second API costs, companies will increasingly fine-tune smaller, proprietary models on local infrastructure for repetitive video tasks.
Timeline
2024-08
MiniMax releases initial video generation capabilities, signaling entry into the generative video market.
2025-03
ByteDance integrates advanced video generation features into its ecosystem, laying the groundwork for Seedance series.
2026-05
MiniMax H3 is unveiled, introducing improved temporal stability and higher resolution outputs.
2026-07
Seedance 2.5 launches, focusing on enhanced physics simulation and motion control.
- 2024-08MiniMax releases initial video generation capabilities, signaling entry into the generative video market.
- 2025-03ByteDance integrates advanced video generation features into its ecosystem, laying the groundwork for Seedance series.
- 2026-05MiniMax H3 is unveiled, introducing improved temporal stability and higher resolution outputs.
- 2026-07Seedance 2.5 launches, focusing on enhanced physics simulation and motion control.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.