CMB International released a research report reviewing the development of internet large models in 3Q26, concluding that the upper bound of model capability remains the key to determining differentiated pricing power, while declining cost per task and agentization are expected to support rapid growth in call volumes.
High value with attractive ROI, verifiable output results, sufficient data, and controllable error costs are the key elements driving further AI penetration in vertical scenarios, and the deepening of RSI may further promote the depth of AI usage.
Review of 3Q26 Industry Development
1) Foundation model intelligence continued to improve, with Claude Opus 5.5 and Sonnet 5.5 leading globally, while Chinese vendors maintained their advantage in the open-source model space, though competition in open-source models also showed signs of intensifying; 2) The industry's breakthrough directions include strengthening computer operation, long-horizon software engineering, and cybersecurity capabilities, pushing coding and agents toward complete workflows; 3) User growth and increased usage depth jointly drove commercialization progress, while domestic large model vendors are expected to further broaden revenue sources through call-volume-based revenue sharing with overseas cloud platforms.
Industry Progress: Computer Operation, Long-Horizon Software Engineering, and Cybersecurity as Key Breakthrough Directions
Real task execution capabilities of models further improved in 3Q26: 1) Computer operation and output editing are gradually being integrated, with products like Cowork connecting data reading, cross-software operations, and document delivery, pushing agents from coding toward general office work; 2) Software engineering capabilities are extending from code generation to long-duration autonomous execution, with models able to continuously develop, test, and fix issues around the same objective; 3) Cybersecurity applications are expanding toward vulnerability detection and automated remediation. CMB International believes that improved long-horizon task completion capabilities will expand the scope of work AI can undertake, providing support for higher-frequency, continuous calls.
Industry Trends: User Base Expansion, Head Paying Groups Increasing Usage Depth
On the user scale front, ChatGPT weekly active users grew from 1 billion in August 2026 to 1.2 billion in October, and the proportion of U.S. consumer samples paying for at least one of ChatGPT, Gemini, or Claude rose from 2.1% in August 2025 to 4.5% in August 2026. On the usage depth dimension, high-intensity B-end users continued to increase investment. In Ramp's U.S. enterprise sample, top 1% enterprises spent $7,205 per employee per month on AI in August 2026, up 198% YoY, while top 10% and median enterprises saw spending growth of 246% and 165% YoY respectively. On C-end paid penetration, spending among top paying groups grew faster. YipitData shows that from January 2025 to August 2026, the average monthly spending of top 1% AI paying users in the U.S. grew 79%, with building and automation, content creation, and transaction processing being important paid scenarios.
Business Model: Overseas Model Commercialization Continues to Deepen, Domestic Model Monetization Channels Further Broadened
Key marginal changes in overseas AI platform commercialization in 3Q26 include: 1) OpenAI's enterprise business growth accelerated, supporting rapid growth in annualized revenue run rate (ARR), with the company's enterprise business ARR growing 107% YoY in 3Q26 (according to CNBC); 2) Advertising commercialization continued to advance. On the domestic AI platform commercialization front, the overseas cloud platform revenue-sharing model was successfully expanded.
Future Outlook: RSI Accelerates Model Iteration, Lower Cost Per Task and Computer Operation Drive AI Penetration Higher, Focus on Agent Progress
CMB International expects that continued deepening of Rush Street Interactive, Inc. (RSI) may further drive AI penetration higher. The proportion of AI-led R&D work disclosed by Anthropic rose to 26% in August (March: 1%). Improved model capabilities and declining cost per task together improve user ROI. Enhanced computer operation capabilities enable AI to participate in more non-coding office processes, and efficiency cases in daily work are expected to form positive demonstration effects, further raising B-end and C-end penetration and usage depth. As agent progress deepens, industry competition is expected to gradually shift from "whose model is stronger" to "who can own users' long-term workflows." The value of middle-layer SaaS and automation tools faces risk of being squeezed, while governance and compliance are expected to receive more attention.
Recommended Targets
1) Overseas cloud vendors: recommend watching Amazon (AMZN.US), Microsoft (MSFT.US), and Google (GOOGL.US); 2) China cloud and AI applications: recommend Alibaba (BABA.US) and Tencent (00700); 3) Large model targets: bullish on Zhipu (02513) and MINIMAX-W (00100).
For more heavyweight Hong Kong stock news, download the Zhitong Finance app. For more Hong Kong and overseas wealth management information, please visit www.zhitongcaijing.com (search "Zhitong Finance"); to join the Zhitong Hong Kong stock investment group, please add Zhitong customer service WeChat (ztcjkf).
Comments