China Securities Co.,Ltd.: Flagship models and Agent products continue to upgrade, and commercial value is expected to be further enhanced.

date
07:45 18/11/2026
avatar
GMT Eight
Flagship models and Agent products continue to upgrade; we remain bullish on leading platforms that combine models, cloud resources, and application entry points.
China Securities Co.,Ltd. released a research report stating that flagship models and Agent products continue to upgrade, and it remains bullish on leading platforms that combine models, cloud resources, and application entry points. Qwen 4 Max and Claude 5.2 strengthen coding, professional tools, and long-task capabilities, Google supplements enterprise inference systems, OpenAI launches merchant services and authorized payments, and AI applications are expanding toward specific business operations. From November 9 to 15, OpenRouter call volume grew 0.86% week-on-week, with Hy4 and MiMo V2.6 Flash growing 41.61% and 87.91% respectively; Claude Code's 7-day average downloads grew 25.37%, and WorkBuddy store revenue grew 62.12%. Model upgrades and improved deployment products are expected to drive enterprises to increase calls and payments, and platforms with leading model capabilities and mature cloud delivery will benefit more fully. China Securities Co.,Ltd.'s main views are as follows: Platform calls grew steadily, with Hy4 and Xiaomi model usage rising From November 9 to 15, OpenRouter's weekly usage was 211.28 trillion tokens, up 0.86% week-on-week. Hy4 usage was 18.12 trillion, up 41.61%, and MiMo V2.6 Flash usage was 4.23 trillion, up 87.91%. Platform totals were stable, while adoption of some domestic models increased relatively quickly. Coding tool adoption increased, with Claude Code download growth leading As of November 14, Claude Code's 7-day average downloads reached 2.52 million per day, up 25.37% from a week earlier; Codex was 3.79 million per day, up 1.88%. Calls to Kilo Code and Codex on OpenRouter also grew, and coding remains a direction with relatively clear Agent demand. WorkBuddy usage and payments both increased, Qwen revenue rose WorkBuddy weekly downloads were 918,900, up 17.91%, store revenue was US$61,200, up 62.12%; Qwen store revenue grew 39.92%. Gemini downloads in the U.S. region grew 24.01%. Application performance diverged, with usage and payment growth in office tools relatively prominent. Flagship models accelerated iteration, with complex task capabilities continuing to strengthen Alibaba released the official version of Qwen 4 Max on November 16, supporting native vision and a 1 million token context; Anthropic released Fable 5.2 and Mythos 5.2. Leading vendors continued to improve coding, professional tools, and long-task capabilities, with model competition penetrating enterprises' actual workflows. Open-source model call share rebounded, with commercialization opportunities covering multiple types of models. On November 15, open-source models' token share on Vercel AI Gateway was 72.57%, up 1.50 percentage points from a week earlier; on the spending side, GLM-5.3's share rose to 12.92%, up 4.88 percentage points, and Kimi K3.1 Pro's share rose 2.10 percentage points. Claude Opus 5.5 remained important on the spending side with a 23.09% share, and the newly released Fable 5.2 had already gained a 5.77% share. Domestic models made progress on both calls and spending, and new closed-source models were also quickly gaining paid demand. Enterprises' model choices became more specific, with cost efficiency and complex task capabilities each having a market, and vendors that continue to iterate and can enter actual workflows have more advantages. Long context and security needs drove continued upgrades to cloud inference systems. Google added TPU 8tx and TPU 8i Max systems: the former is based on TPU 8t and combines fifth-generation AMD EPYC processors to provide hardware protection for models and data during inference, targeting sensitive businesses such as finance and government; the latter integrates two TPU 8i chips in the same package, increasing compute, on-chip storage, and memory bandwidth to support ultra-long context and multi-Agent tasks. Model capability improvements are bringing more complex deployment requirements, and cloud platforms are launching specialized products around security, latency, and long-task processing, which helps capture demand as enterprises move from trials to production deployment. Platforms with model services, self-developed compute, and an enterprise customer base are expected to expand their cloud businesses in this process. GPU rental prices diverged, with high-end inference resources still supported by demand. As of November 14, the H100 rental price was US$3.19/GPU-hour, up 2.90% from a week earlier; H200 was US$6.60/GPU-hour, up 0.46%, and rose to US$7.75/GPU-hour at one point during the week. B200 fell to US$9.66/GPU-hour, down 2.23%, A100 fell 1.92%, and RTX 5090 was flat. Prices for different models did not rise in sync, while gains in H100 and H200 indicated that demand for some high-end resources remained resilient. Combined with Google's new systems for long context and confidential inference, compute products are being further segmented around specific tasks, and cloud service providers that can offer suitable resources and improve deployment efficiency are better positioned to capture incremental demand. Merchant services and authorized payments completed the transaction link, and Agent applications expanded into commercial scenarios. OpenAI launched a beta version of ChatGPT merchant services, through which merchants can maintain profiles, prices, and availability via Business Hub, and eligible U.S. merchants can also configure Business Agent to answer product and service questions; ChatGPT Wallet is gradually opening to eligible U.S. adult Pro users and supports credit card payments after user authorization. The integration of merchant information, product Q&A, and payment functions enables ChatGPT to handle more transaction-related tasks and also provides a direction for platforms with user entry points to expand merchant services. As tool calling and long-task capabilities strengthen, the scope of services Agents can provide is expanding from content generation to specific business operations, and the commercial value of application entry points is expected to increase further. Continue to favor leading platforms where model capabilities, cloud services, and application entry points create synergies. In this week's high-frequency data, total platform calls grew slightly, while coding tool downloads and payments for some office applications grew more prominently; on the product side, Qwen and Claude continued to improve complex task capabilities, Google supplemented enterprise inference systems, and OpenAI extended merchant and payment functions. Demand growth and product upgrades are reinforcing each other in specific scenarios. On investment, focus on platforms that can continuously launch high-capability models, capture enterprise calls through cloud services, and rely on application entry points to expand payment scenarios; among them, coding and office Agents already have relatively clear usage and payment progress and are important directions for observing AI commercialization at present. Risk warnings Risks of technological iteration and intensified competition; risk that commercialization monetization falls short of expectations; risks related to compute delivery, costs, and return on capital; content security, payment security, and regulatory risks; and risks of high-frequency data tracking and estimation deviations.