China's AI model cost advantage continues, with the Hy3 official release and Token trends as key focus areas

Overview

This report is the 55th edition of Jefferies' China Internet AI series, covering the Hy3 official release, Token consumption trends, model input/output pricing, video generation models, AI assistants, monthly active users of AI mobile applications, cloud service monetization, Agent platform heat maps, and Chinese internet application traffic. Using a China-US AI model comparison as its central theme, the report emphasizes the progress of Chinese models in cost efficiency, usage, and expansion of application scenarios.

Core views

The core views include: first, the Hy3 official version improves over the preview version in task success rate, cost efficiency, long-context capability, and hallucination control, while supporting broader users and scenarios with lower API prices; second, OpenRouter data shows that Chinese model Token consumption exceeded that of US models in the latest week, with DeepSeek V4 Flash, MiMo V2.5, and MiniMax M3 ranking near the top; third, although the Token expenditure index remains soft, AI Coding, Agents, and high-performance, low-cost models are driving medium-term demand; fourth, the intelligence gap between Chinese and US models is narrowing, while API costs are only a fraction of those of US models; fifth, regulators have introduced measures targeting anthropomorphic virtual intimate relationship services, which may affect certain AI companion functions.

Analysis framework

The report combines OpenRouter Token consumption data, the Silicon Data LLM Token Expenditure Index, Artificial Analysis model intelligence and pricing data, QuestMobile and aicpb.com application traffic data, as well as third-party evaluations and industry news, to track AI model supply and demand, price competition, application traffic, and cloud service monetization.

Methodology notes

• Usage trackingOpenRouter Token consumption analysis

  Weekly Token consumption statistics by model and company

  Compares the latest week with the prior week to observe Token usage, model rankings, and changes in company shares for Chinese and US models.

• Spending intensity indicatorSilicon Data LLM Token Expenditure Index

  Weighted Token expenditure index

  The index is weighted by usage concentration; it rises when demand shifts toward high-end models and is used to assess Token spending strength.

• Model evaluationArtificial Analysis

  Comparison of intelligence index, pricing, speed, and cost efficiency

  Compares the competitiveness of Chinese and US foundation models across intelligence, API pricing, output speed, and cost.

• Application traffic trackingQuestMobile and aicpb.com application data

  DAU, MAU, and month-over-month and year-over-year changes

  Used to track changes in user scale for AI assistants, AI mobile applications, and applications in Chinese internet subsectors.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Baidu、Alibaba、Tencent

  Potential beneficiaries of China's cloud services and AI model ecosystem

  Strengths

  Possess cloud resources, model capabilities, application entry points, and an enterprise customer base.

  Weaknesses

  The pace of monetization is affected by model costs, corporate budgets, and regulatory boundaries.

  Comparison

  Chinese model API costs are lower than those of US models, while the intelligence gap is narrowing.

  Risks

  Soft Token demand, intensifying price competition, and regulatory restrictions on anthropomorphic services.

• Kingsoft Cloud

  Potential beneficiary of rising AI cloud demand

  Strengths

  May benefit from cloud computing and service demand generated by AI models and Agents.

  Weaknesses

  Sensitive to the intensity of industry AI spending and customer budgets.

  Comparison

  May be weaker in scale and ecosystem coverage than large CSPs.

  Risks

  Cloud service pricing pressure and uncertain demand elasticity after free quotas are removed.

• DeepSeek、Qwen、Kimi、MiniMax、Zhipu

  Representatives of improving Chinese foundation model competitiveness

  Strengths

  High cost efficiency, with some models ranking near the top in OpenRouter usage and third-party evaluations.

  Weaknesses

  Long-term monetization, ecosystem development, and enterprise-level deployment capabilities still require continued validation.

  Comparison

  Lower prices than US models, with a narrowing intelligence gap.

  Risks

  Changes in pricing strategy, computing constraints, and model iterations falling short of expectations.

• Kling and video generation model platforms

  A beneficiary direction of growing content and video generation demand

  Strengths

  Upgrades in video generation, editing, and long-form video capabilities can improve content production efficiency.

  Weaknesses

  Monetization is affected by subscription prices, generation quality, and user retention.

  Comparison

  Competes with models such as Seedance, Wan, and SkyReels.

  Risks

  Model quality competition, copyright and content regulatory risks, and pressure on unit generation costs.

Key data

• Hy3 official version API pricingInput RMB1 per million Tokens, output RMB4 per million TokensBelow the preview version's input price of RMB2 per million Tokens and output price of RMB8 per million Tokens.
• Hy3 task success rate90%Up from 72% for the preview version; the report also mentions a 15-percentage-point reduction in the hallucination rate.
• OpenRouter weekly Token consumption46.7tnStable week over week during the week of June 29.
• Chinese model Token consumption23.4tn, up 14.9% week over weekExceeded US model consumption by 4.3tn during the week of June 29 to July 5.
• Model Token consumption rankingsDeepSeek V4 Flash 5.34tn; MiMo V2.5 4.38tn; MiniMax M3 4.11tnDeepSeek V4 Flash ranked first in the latest week.
• Company Token consumption shareGoogle 28.2%; DeepSeek 20.5%; OpenAI 16.2%; Qwen 5.8%; Anthropic 5.2%Weekly company shares of Token consumption as of July 6.
• Token Expenditure Index1.62 to 1.71The range from June 28 to July 4, below 2.04 on May 31.
• AI assistant DAUDoubao 158m; DeepSeek approximately 30.5m; Qwen 27.9m; Yuanbao 8.8mMay data; Doubao increased approximately 5% month over month, while DeepSeek increased approximately 6%.

Impact & implications

The report is positive on China's AI infrastructure and application ecosystem. If demand for AI models and Agents continues to rise while supply-side computing and cost constraints remain, low-cost, high-performance models, cloud service providers, and AI application platforms will become more competitive. Tencent, Alibaba, Baidu, Kingsoft Cloud, AI labs, and Kling are viewed as potential beneficiaries; however, regulatory constraints on AI companion services, corporate controls on Token usage, and model price changes could affect the pace of short-term monetization.

Risks

• Enterprises and internet companies setting limits on employee Token consumption could suppress near-term spending intensity.
• Model API price competition or peak-period price increases could change developers' usage patterns and demand elasticity.
• Regulatory measures targeting anthropomorphic AI services could affect virtual companions, virtual intimate relationships, and related functions.
• Although Chinese models have a cost advantage, their intelligence, ecosystems, and depth of commercialization relative to leading global models still require continued validation.
• Removing free cloud service quotas could promote monetization but may also reduce some trial and long-tail demand.

What to watch

• Adoption, retention, and paid conversion of the Hy3 official version in real-world work scenarios.
• Changes in Token consumption after DeepSeek V4's peak-period price adjustment starting July 15.
• Market feedback on Seedance 2.5's video generation quality, duration, and multimodal input capabilities after its release.
• Whether Chinese models' weekly Token share relative to US models in OpenRouter continues to expand.
• Whether the Silicon Data LLM Token Expenditure Index can recover to late-May levels.
• The impact of product function adjustments by Doubao, Qwen, and others on user metrics after anthropomorphic service regulations take effect in mid-July.