๐ญ๐ฐSCMP TechnologyโขRecentcollected in 2m
China's AI Price War Intensifies Among Major Tech Giants
๐กMajor Chinese tech firms are slashing AI prices; learn how this shift impacts your infrastructure and inference costs.
โก 30-Second TL;DR
What Changed
ByteDance and Tencent have launched aggressive AI pricing offensives.
Why It Matters
The price war will likely accelerate AI adoption in China but may squeeze profit margins for smaller AI startups. Developers can expect lower inference costs as competition forces providers to optimize efficiency.
What To Do Next
Evaluate your current LLM inference costs and compare them against the latest pricing tiers from ByteDance and Tencent to optimize your infrastructure spend.
Who should care:Founders & Product Leaders
๐ง Deep Insight
Web-grounded analysis with 33 cited sources.
๐ Enhanced Key Takeaways
- โขThe AI price war in China extends beyond ByteDance and Tencent, actively involving other major players such as Alibaba, Baidu, DeepSeek, and Xiaomi, with DeepSeek often initiating significant price cuts that compel rivals to follow suit.
- โขChinese tech giants have implemented drastic price reductions, with some models seeing cuts of up to 99%, making their AI APIs substantially more affordable than many Western counterparts. For example, ByteDance cut Doubao Pro-32K by 99.3% in May 2024, and Xiaomi slashed MiMo-V2.5 API costs by up to 99% in May 2026.
- โขBeyond simply capturing market share, the aggressive pricing strategies aim to democratize AI access for small and medium-sized businesses, foster a robust developer ecosystem, and strategically shift AI capabilities from 'technology demonstration' to 'value monetization' for specific model tiers.
- โขWhile many models are seeing price reductions, Tencent Cloud notably implemented significant price increases (over 450%) for some of its Hunyuan series models, such as HY2.0 Instruct, in March 2026, indicating a strategic move towards commercialization and confidence in the value of its advanced offerings.
- โขThe Chinese government has signaled a desire to curb 'involution-style' competition, which refers to aggressive price wars and subsidies, and instead encourages tech platforms to increase investment in strategic AI technologies and compete on value, suggesting a potential shift in regulatory focus.
๐ Competitor Analysisโธ Show
Chinese AI Model Comparison (as of June 2026)
| Company/Model | Key Features | Pricing (per 1M output tokens) | Benchmarks/Performance Highlights |
|---|---|---|---|
| ByteDance Doubao Pro 256K | 256k context length, full-modality support, optimized for Chinese cultural reinforcement, vertical scenario optimization. | Lowest listed input price for Doubao Lite 32K at $0.044/1M input tokens. Seedance 2.0 Mini at ~$0.073/second for video. | Strong in math-heavy prompts (Doubao), multimodal video generation (Seedance). |
| Tencent Hunyuan A13B Instruct | Integrated with WeChat/QQ ecosystem, strong Chinese language support, enterprise solutions, gaming AI, cloud integration, 131k context length. | $0.71 (input + output combined). Hunyuan HY2.0 Instruct surged to ยฅ0.004505 per thousand tokens (approx. $0.62/1M tokens) in March 2026. | Competitive performance across math, logic, coding, science, and agentic tasks. |
| Alibaba Qwen3 Max | Multilingual master, multimodal capabilities (Qwen-VL-Max for visual reasoning), MoE architecture. | $3.90. Qwen-VL-Max at ยฅ0.003 per thousand tokens (approx. $0.41/1M tokens). | Comparable to GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3. Strongest for multilingual applications (Chinese, Japanese, Korean). |
| Baidu ERNIE X1 / 4.5 | ERNIE X1: Deep reasoning, multimodal capabilities, tool-use. ERNIE 4.5: Multimodal, high emotional intelligence. | ERNIE X1: $1.10. ERNIE 4.5: $2.20. ERNIE 4.5 21B A3B Thinking: $0.35 (input + output combined). | ERNIE X1 comparable to DeepSeek's R1 reasoning model. ERNIE 5.1 ranked first among Chinese models on Arena benchmark. |
| DeepSeek V4-Pro | Enhanced MoE architecture, excels in coding and mathematical tasks, low inference costs. | $0.87. DeepSeek V3.2 at $0.28/1M input tokens. | Leads BenchLM's Chinese leaderboard (V4 Pro Max at 87). Outperforms other open-source models and rivals leading closed-source models in coding. |
| Xiaomi MiMo-V2.5 Pro | Multimodal, long-context processing. | $3.00 (flat across input length). | Competitive performance. |
| Zhipu AI GLM-5 | 744B parameter MoE (40B active), strong in agentic/coding benchmarks, low hallucination rates. | $3.20 | Leads overall Chinese rankings with BenchLM score of 85, 77.8% on SWE-bench Verified. Approaches Claude Opus 4.5 in agentic/coding. |
| Moonshot AI Kimi K2.6 | Advanced multimodal integration, long-context processing (up to 1M tokens), agentic tasks. | Lowest cache-hit floor at $0.07. | Dominates agentic benchmarks (76.8% SWE-bench, 74.9% BrowseComp). |
๐ ๏ธ Technical Deep Dive
- ByteDance Doubao/Seedance Models: Utilize an advanced Mixture-of-Experts (MoE) architecture. They offer full-modality support, covering text, images, speech, video, and 3D, with a focus on vertical scenario optimization and Chinese cultural reinforcement. Doubao Pro supports a 256,000 token context length, while Doubao Lite offers 128,000 tokens. Seedance 2.0 Mini is designed for everyday production workflows, prioritizing speed and repeated experimentation.
- Tencent Hunyuan Models: Built on a Transformer-based Mixture-of-Experts (MoE) architecture. The Hunyuan-Large model features a total of 389 billion parameters with 52 billion active parameters. It supports a pre-trained context length of up to 256,000 tokens and an Instruct model context length of up to 128,000 tokens. Key technical innovations include high-quality synthetic data for richer representations and long-context handling, KV Cache Compression using Grouped Query Attention (GQA) and Cross-Layer Attention (CLA) to reduce memory and improve inference throughput, and Expert-Specific Learning Rate Scaling for optimized learning. Tencent Hunyuan also integrates with NVIDIA TensorRT-LLM for high-performance inference.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
The intense price competition will accelerate AI adoption across various industries in China, particularly among small and medium-sized enterprises (SMEs).
Lower costs significantly reduce the financial barriers to entry for advanced AI technologies, making them accessible to a broader range of businesses that previously found them cost-prohibitive.
Chinese AI models are poised to gain significant market share globally, especially in price-sensitive developing countries.
The drastic price reductions make Chinese models highly competitive, and with their rapidly converging capabilities compared to Western counterparts, they offer an attractive, cost-effective alternative in international markets.
The focus of competition among Chinese tech giants will gradually shift from raw price cuts to differentiation through specialized models, value-added services, and operational efficiency.
Signals from the Chinese government to curb 'involution-style' competition, combined with the need for sustainable business models, will push companies to innovate beyond just price, emphasizing quality, unique features, and efficient resource utilization.
โณ Timeline
2024-05
ByteDance initiates major price cut for Doubao Pro-32K by 99.3%, triggering broader market response.
2024-11
Tencent open-sources Hunyuan-Large MoE model with 389 billion parameters.
2025-01
Alibaba Cloud slashes Qwen-VL-Max prices by 85%, intensifying competition.
2025-03
Baidu launches ERNIE 4.5 and ERNIE X1 with aggressive pricing, undercutting DeepSeek.
2025-11
Alibaba further reduces charges for its Qwen3-Max model by up to 50%.
2026-05
DeepSeek makes V4 discount permanent; Xiaomi cuts MiMo-V2.5 API costs by up to 99%. Chinese Communist Party journal signals to tech giants to curb price wars.
๐ Sources (33)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- medium.com
- cloudcomputing-news.net
- techtimes.com
- warontherocks.com
- scmp.com
- trendingtopics.eu
- biggo.com
- thenextweb.com
- llmreference.com
- aijourn.com
- crazyrouter.com
- aipricing.org
- youtube.com
- alphamatch.ai
- zenmux.ai
- dev.to
- remoteopenclaw.com
- aimagazine.com
- techtarget.com
- aipricing.org
- techinasia.com
- index.dev
- benchlm.ai
- howaiworks.ai
- substack.com
- reddit.com
- arxiv.org
- huggingface.co
- github.com
- nvidia.com
- channelnewsasia.com
- rand.org
- scmp.com
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