๐ญ๐ฐSCMP TechnologyโขStalecollected in 12h
DeepSeek V4 triggers aggressive AI price war in China
๐กDeepSeek's aggressive pricing is reshaping the Chinese AI market; see how competitors like Xiaomi are responding.
โก 30-Second TL;DR
What Changed
DeepSeek V4 pricing model is forcing competitors to overhaul monetization strategies.
Why It Matters
This price war significantly lowers the barrier to entry for developers building on Chinese LLMs. It may force smaller AI startups to pivot their business models away from pure API reselling.
What To Do Next
Evaluate your current LLM spend and test DeepSeek or MiMo-V2.5 APIs to see if you can reduce your inference costs by up to 99%.
Who should care:Developers & AI Engineers
Key Points
- โขDeepSeek V4 pricing model is forcing competitors to overhaul monetization strategies.
- โขXiaomi reduced MiMo-V2.5 API costs by 99% to remain competitive.
- โขThe Chinese AI market is shifting toward a cutthroat pricing environment for cloud-based inference.
๐ง Deep Insight
Web-grounded analysis with 20 cited sources.
๐ Enhanced Key Takeaways
- โขDeepSeek V4's pricing for its V4-Flash model is as low as $0.14 per million input tokens and $0.28 per million output tokens, making it significantly cheaper than Western counterparts like GPT-5.5 and Claude Opus 4.8, which can be 10-50 times more expensive.
- โขThe aggressive pricing by DeepSeek V4 has led to a surge in usage for competing Chinese models, with Xiaomi's MiMo-V2.5 processing 1.7 trillion tokens in a week, representing over 999% growth, and climbing to sixth place on the OpenRouter marketplace.
- โขDeepSeek V4 models, including V4-Pro and V4-Flash, are open-weight and released under the MIT license, enabling broader adoption and self-hosting, a strategy that contrasts with the proprietary 'black box' approach of many Western AI companies.
- โขThe Chinese AI price war, initiated by models like DeepSeek, has caused a structural shift in the global AI landscape, collapsing costs for equivalent intelligence by 90-97% and challenging the assumption that frontier AI requires billions of dollars in training.
- โขDeepSeek V4-Pro, with 1.6 trillion total parameters and 49 billion active parameters, achieves high performance on coding benchmarks, scoring 80.6% on SWE-bench Verified and 93.5 on LiveCodeBench, while also offering native multimodal capabilities.
๐ Competitor Analysisโธ Show
| Feature/Model | DeepSeek V4 Pro | DeepSeek V4 Flash | Xiaomi MiMo-V2.5 | OpenAI GPT-5.5 (approx.) | Anthropic Claude Opus 4.7/4.8 (approx.) |
|---|---|---|---|---|---|
| Total Parameters | 1.6 Trillion (MoE) | 284 Billion (MoE) | N/A (MiMo-V2-Pro > 1T) | N/A | N/A |
| Active Parameters | ~49 Billion/token | ~13 Billion/token | N/A | N/A | N/A |
| Context Window | 1 Million tokens | 1 Million tokens | 1 Million tokens | N/A | N/A |
| Multimodal | Native (text, images, video, audio) | N/A (text/code focus) | Native Omnimodal (text, images, video, audio) | N/A | N/A |
| Input Pricing (per 1M tokens) | $0.435 | $0.14 | $0.14 | $5.00 | $5.00 |
| Output Pricing (per 1M tokens) | $0.87 | $0.28 | $0.28 | $30.00 | N/A |
| Key Benchmarks | SWE-bench Verified: 80.6%, LiveCodeBench: 93.5 | N/A (trails Pro by 7-10 points on agentic coding) | Artificial Analysis Intelligence Index: 49 | N/A | N/A |
๐ ๏ธ Technical Deep Dive
- DeepSeek V4 Architecture: Utilizes a Mixture-of-Experts (MoE) design, with V4-Pro having 1.6 trillion total parameters and ~49 billion active per token, and V4-Flash having 284 billion total parameters and ~13 billion active per token.
- Hybrid Attention Architecture: DeepSeek V4 introduces a novel Hybrid Attention Architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to significantly enhance efficiency for long context windows.
- Manifold-Constrained Hyper-Connections (mHC): This innovation constrains residual mapping to improve signal propagation stability while maintaining model expressivity.
- Muon Optimizer: DeepSeek V4 employs the Muon optimizer, which contributes to faster convergence and improved training stability.
- Context Window: Both DeepSeek V4 Pro and Flash models support an extensive 1 million token context window.
- Multimodality: DeepSeek V4 (Base and Pro) is natively multimodal, trained from scratch on text, images, video, and audio simultaneously.
- Xiaomi MiMo-V2.5: Described as a native omnimodal model, capable of processing text, images, video, and audio within a single architecture, and features a 1 million token context window.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
The aggressive AI pricing in China will accelerate global AI adoption and shift market share towards more cost-efficient models.
Drastically lower API costs make advanced AI accessible to a wider range of developers and businesses, particularly in cost-sensitive markets, fostering new applications and increasing overall token consumption.
Western AI companies will be forced to further reduce their API pricing and potentially adopt more open-source strategies to remain competitive.
The significant price disparity and comparable performance of Chinese models are creating immense pressure on Western providers, who risk losing market share if they maintain premium pricing.
The focus on efficiency and open-source models by Chinese AI labs will drive innovation in AI architecture and hardware optimization.
Operating under hardware constraints and a competitive pricing environment, Chinese companies have developed efficient architectures like MoE and novel attention mechanisms, pushing the boundaries of cost-effective AI.
โณ Timeline
2023-07
DeepSeek founded by Liang Wenfeng, funded by High-Flyer Capital.
2023-11
DeepSeek Coder and DeepSeek-LLM series released.
2024-05
DeepSeek-V2, an MoE model with 236B parameters and 128K context, released.
2024-12
DeepSeek-V3-Base and DeepSeek-V3 (chat) released.
2025-01
DeepSeek-R1 model and chatbot launched, gaining international prominence and becoming a top downloaded app.
2026-04
DeepSeek V4 (Base, Pro, Flash) models released, featuring 1M token context and multimodal capabilities.
๐ Sources (20)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology โ