Chinese AI models stand out for their cost advantage, while Hy3's official release reinforces model efficiency and application scenarios

Overview

This is the 55th report in Jefferies' China Internet AI Series, covering the official release of Hy3, token consumption trends, model pricing, OpenRouter usage, AI application user data, video generation models and regulatory changes. The report compares Chinese and US models, concluding that while the intelligence gap is narrowing, Chinese models have a significant advantage in API costs due to architectural and computing-efficiency advantages including MoE, attention mechanisms and MLU.

Core views

The core views include: first, the official version of Hy3 improves over the preview version in user experience, task success rate, token efficiency and hallucination control, while also offering lower pricing, which should help expand user and use-case coverage; second, the Token Expenditure Index remained soft during the week of 28 Jun and below May levels, although overall OpenRouter token usage was stable; third, Chinese models' token consumption on OpenRouter is growing rapidly and exceeded that of US models in the latest week; fourth, Coding, Workspace, long- and medium-form content creation and AI Agents represent important opportunities; fifth, as AI demand rises and supply constraints persist, cost-effective models and cloud service infrastructure will become more strategically valuable.

Analysis framework

The report uses cross-validation across multiple data sources, combining OpenRouter token consumption, the Silicon Data LLM Token Expenditure Index, Artificial Analysis model intelligence and pricing metrics, QuestMobile application user data, PinchBench success-rate rankings and the Omdia agentic AI platform heatmap to track model competitiveness, pricing trends, user demand, application scenarios and potential industry-chain beneficiaries.

Methodology notes

• Usage TrackingOpenRouter Token Consumption Analysis

  Measures model usage popularity through token consumption by different models and companies on OpenRouter.

  The report compares token consumption during the weeks of 29 Jun and 22 Jun, and observes the relative positions of DeepSeek, Xiaomi, MiniMax, Qwen, OpenAI, Google and other players by model and company rankings.

• Pricing and Demand IntensitySilicon Data LLM Token Expenditure Index

  Measures AI model demand and shifts between model tiers through a token expenditure index weighted by usage mix.

  The index ranged from 1.62 to 1.71 between 28 Jun and 4 Jul, below the 1.68 to 1.74 range from 21 Jun to 27 Jun and also below 2.04 on 31 May, indicating that short-term spending remained soft.

• Model CompetitivenessArtificial Analysis

  Compares model intelligence, agentic capabilities, coding, output speed and API costs.

  The report uses this framework to conclude that the intelligence gap between Chinese and US models is narrowing, while emphasizing the cost advantage of Chinese models in input/output API pricing.

• User BehaviorQuestMobile App Tracker

  Tracks DAU, MAU and usage duration for Chinese internet and AI applications.

  The report cites May-26 data to compare user trends for Doubao, Qwen, Yuanbao and DeepSeek, as well as applications in e-commerce, music, video, short video, travel and other internet subsectors.

• Platform EcosystemOmdia agentic AI development platform heatmap

  Compares cloud platforms across model support, multi-agent frameworks, lifecycle management, tool marketplaces, vertical solutions and open-source communities.

  The report presents an agentic AI development platform heatmap for Asia and Oceania in its appendix, covering Alibaba Cloud, AWS, Google Cloud Platform, Microsoft Azure and Tencent Cloud.

Asset mapping & comparison

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

• Baidu, Alibaba, Tencent

  Potential Beneficiaries

  Strengths

  Broad coverage across cloud services, models, application entry points and AI ecosystems; the report believes rising token consumption will benefit CSPs like Baidu, Alibaba and Tencent.

  Weaknesses

  The report does not provide company-specific financial elasticity, valuation sensitivity or margin impacts.

  Comparison

  Compared with single-model companies, BAT has advantages from synergies across cloud, applications and ecosystems.

  Risks

  Model price competition, computing-supply constraints, enterprise customer controls on token spending and regulatory changes.

• Kingsoft Cloud

  Potential Beneficiary

  Strengths

  As a cloud service provider, it may benefit from rising token consumption driven by demand for AI models and Agents.

  Weaknesses

  The report does not disclose specific orders, revenue contributions or margin data.

  Comparison

  Compared with large BAT cloud platforms, its scale and ecosystem coverage may be weaker, but it remains exposed to expanding AI cloud demand.

  Risks

  Cloud price competition, capital expenditure pressure and fluctuations in AI demand.

• DeepSeek, Qwen, Kimi, MiniMax, Zhipu

  Representative Chinese Large Models

  Strengths

  The report believes these Chinese models are catching up with US models in intelligence and have significant advantages in API costs.

  Weaknesses

  Some models still face uncertainty regarding commercialization, ongoing training investment and ecosystem expansion.

  Comparison

  Compared with US models, the API costs of Chinese models are described as only a fraction of those of US models.

  Risks

  Price wars, user acceptance of peak-hour price increases, failed technology iterations and regulatory constraints.

• Tencent Hy3

  Key Model to Track

  Strengths

  Compared with the preview version, the official version improves success rate, token efficiency, hallucination rate and long-context tasks, while lowering prices.

  Weaknesses

  The report does not provide actual commercial revenue or customer retention data following the official release of Hy3.

  Comparison

  Blind tests show that Hy3 outperforms GLM-5.1 in certain real-world work scenarios and is more cost-efficient than the preview version.

  Risks

  Intensifying model competition, user switching costs and real-world performance falling below evaluation results.

• OpenAI, Anthropic, Google

  US Model Comparison Group

  Strengths

  They remain in the global top tier in Arena Elo, success rates and certain intelligence evaluations.

  Weaknesses

  The report emphasizes that their API costs are higher than those of Chinese models.

  Comparison

  US models lead in some capability metrics, but the intelligence gap is narrowing and Chinese models have a clear cost advantage.

  Risks

  If Chinese models continue improving their value for money, US models may face share pressure in price-sensitive scenarios.

• Kling and the Video Generation Model Ecosystem

  AI Content Creation Opportunity

  Strengths

  The report expects the Seedance 2.5 video generation model to improve in 30-second video generation, multimodal input, video editing and generation quality, expanding opportunities in content creation.

  Weaknesses

  The business model, subscription conversion and unit economics of video generation models remain to be validated.

  Comparison

  The report uses video generation model pricing, subscription plans and ranking performance as horizontal comparison dimensions.

  Risks

  High model costs, copyright and content compliance issues and insufficient user willingness to pay.

Key data

• Hy3 Official Version PricingInput RMB1/million tokens, output RMB4/million tokensBelow the preview version's RMB2/RMB8, supporting more users and use cases.
• Hy3 Task Success Rate90%Up from 72% for the preview version; the report also mentions a 15-percentage-point reduction in hallucination rate.
• Token Expenditure Index1.62 to 1.71For the period from 28 Jun to 4 Jul, below 2.04 on 31 May.
• Weekly OpenRouter Token Consumption46.7tnBroadly stable compared with the previous week during the week of 29 Jun.
• OpenRouter Token Consumption by Chinese Models23.4tn, up 14.9% week over weekFor the period from 29 Jun to 5 Jul, higher than the 4.3tn consumed by US models.
• Company Token Consumption SharesGoogle 28.2%, DeepSeek 20.5%, OpenAI 16.2%, Qwen 5.8%, Anthropic 5.2%Weekly company ranking by token consumption as of 6 Jul.
• Model Token Consumption RankingDeepSeek V4 Flash 5.34tn, MiMo V2.5 4.38tn, MiniMax M3 4.11tnThe next two after the top three were Hy3 preview at 3.13tn and GLM 5.2 at 2.58tn.
• Chinese AI Assistant DAUDoubao 158m, DeepSeek approximately 30.5m, Qwen 27.9m, Yuanbao 8.8mMay-26 data; Doubao increased approximately 5% month over month and DeepSeek approximately 6%.
• Top Three Models by OpenClaw Token ConsumptionMiniMax M3, DeepSeek V4 Flash, DeepSeek V4 ProThe report states that the leading models by token consumption on OpenClaw are mainly from China.
• Chinese Market Model Success RatesQwen3.7 Max 92.5%, MiMo V2.5 89.7%, Qwen3.6 Flash 88.1%, MiMo V2.5 Pro 87.5%From PinchBench, as of 6 Jul.

Impact & implications

The report implies that competition among Chinese AI models is shifting from simple capability comparisons toward comprehensive competition across capability, cost, usability, application scenarios and regulatory constraints. If token consumption continues to expand in Coding Agent, Workspace Agent, video generation and enterprise productivity scenarios, platforms with low-cost models, cloud infrastructure, AI application entry points and developer ecosystems may capture greater value across the industry chain. At the same time, soft short-term token expenditure, corporate restrictions on employee token usage, model price competition and regulatory constraints may affect the pace of commercialization.

Risks

• The Token Expenditure Index remained soft during the week of 28 Jun and below May levels, indicating that short-term AI model spending intensity may be slowing.
• Media reports indicate that technology and internet companies are imposing caps on employee token consumption, which may suppress the release of internal enterprise demand.
• Zhipu and MiniMax face lock-up expiry, which may cause volatility in related assets or sentiment.
• Regulatory measures concerning anthropomorphic services will take effect on 15 Jul, potentially affecting virtual intimate relationships, virtual companionship and other AI application features.
• Model API price competition and peak-hour price increases may coexist, potentially affecting user switching, margins and demand stability.
• Rising AI demand and computing-supply constraints may lead to volatility in service quality, costs and the pace of commercialization.
• User data in certain internet subsectors has weakened, such as year-over-year declines in MAU for long-form video platforms and TCOM/Qunar.

What to watch

• Actual API usage, customer expansion and use-case adoption following the official release of Hy3.
• Whether weekly OpenRouter token consumption remains stable or resumes growth, particularly changes in the share of Chinese models relative to US models.
• Whether the Silicon Data LLM Token Expenditure Index rebounds from its late-June low.
• User acceptance and changes in token usage after DeepSeek V4 doubles peak-hour input/output prices from 15 Jul.
• Performance of Seedance 2.5 in video generation quality, efficiency and user demand following its release.
• Paid conversion after the commercialization of services such as AliCloud Databridge Agent and Tencent Cloud AI-native GPU development and deployment from 1 Aug.
• Month-over-month trends in DAU/MAU for Chinese AI assistants including Doubao, Qwen, Yuanbao and DeepSeek.
• The impact of formal implementation of anthropomorphic service regulations on AI companionship features and application retention.
• The pace at which market scale materializes in Coding, Workspace and medium- to long-form content creation.