Sinolink Securities has released a research report stating that the launch of GPT-5.4 signifies a critical transition for AI from "conversational generation" to "task execution." Through upgrades such as native Computer Use, a Tool Search mechanism, and ultra-long context capabilities, the model can now perform complex operations in real-world environments. The open-source framework OpenClaw is highly complementary to it. The combination of the two significantly reduces the technical costs of AI Agents, enhances multi-tool invocation efficiency, and improves task execution capabilities, thereby advancing AI from an experimental tool to a genuine productivity system.

The main points from Sinolink Securities are as follows:

GPT-5.4 Release: A Major Upgrade Towards "Executable AI" The launch of GPT-5.4 marks a significant leap in large model capabilities from "dialogue generation" to "task execution." This upgrade systematically strengthens reasoning abilities, programming skills, and agent workflow management. For the first time, it integrates Computer Use, tool invocation, and web search capabilities on a large scale at the model level, enabling AI to accomplish complex tasks in real software environments. Specifically: 1) Native Computer Use allows the model to visually understand screens and perform operations like mouse clicks and keyboard inputs. In benchmarks such as OSWorld, its success rate approaches human levels, indicating AI's ability to execute tasks in real systems. 2) The Tool Search mechanism significantly optimizes the tool invocation process by retrieving tool definitions on-demand, reducing context token consumption, thereby improving the efficiency of multi-tool Agent systems and lowering costs. 3) Enhanced web search and reasoning capabilities enable the model to integrate information from across the internet through multi-step searches to solve complex problems. 4) Knowledge-based work capabilities have improved markedly, with performance in professional office tasks—such as spreadsheets, documents, and presentations—approaching or even surpassing that of some industry professionals.

Overall, GPT-5.4 validates two major trends in AI development: first, the accelerating pace of model iteration, and second, the critical role of Computer Use and tool invocation in bridging the gap from "dialogue" to "execution," making the commercialization path for Agents increasingly clear.

OpenClaw + GPT-5.4: An Ideal Combination for the Agent Era As the AI industry enters the Agent era, market focus is shifting from individual model capabilities to the systemic power of "model + Agent framework." In this context, the combination of the open-source Agent framework OpenClaw and GPT-5.4 is regarded by developer communities and media as one of the most promising technical pathways currently available.

OpenClaw, a fully open-source and locally deployable AI Agent framework, aims to enable AI to directly operate computers for real-world work tasks. GPT-5.4's upgrades in several key areas make it highly complementary to OpenClaw: 1) Native Computer Use allows the model to directly operate desktop software, reducing the technical costs associated with complex adaptations and automation scripts in traditional Agent frameworks. 2) An ultra-long context of 1 million tokens provides long-running Agents with greater "working memory," maintaining stable contextual understanding during complex task chains and extended execution periods. 3) The Tool Search mechanism significantly cuts token costs from multi-tool invocations, allowing OpenClaw to maintain high efficiency even when integrating numerous Skills and tools. 4) Enhanced model reasoning and professional task capabilities enable Agents to handle more complex, higher-value knowledge work scenarios.

Overall, the integration of OpenClaw and GPT-5.4 is expected to drive AI Agents from experimental tools toward genuine productivity systems, accelerating the adoption of AI in office automation, software development, and complex knowledge work.

Investment Recommendations Domestic Computing Power: Cambricon, Dongyangguang, Hygon, Syneron, YUNENG, Huafeng Technology, E-TON, Transwarp, Wangsu, Hongjing Technology, Capital Online, Digital China, Baidu, Dawei Technology, Runjian, SMIC, Huahong, Sugon, Runtek, Inspur, Dongshan Precision, Aofei Data, YunSai, Ruicheng, Kehua Data, Weichai Heavy, Kingsoft Cloud, Oulutong, Jiechuang Intelligent.

Overseas Computing/Storage: Zhongji Innolight, Eoptolink, GigaDevice, Dapu Micro, TFC, Yuanjie, Shenghong, Jingwang, Envic, etc.; Lumentum, SanDisk, Kioxia, Micron, SK Hynix, AMEC, Naura, Piotech, Changchuan.

CPU: Hygon, Sugon, Montage, Hesheng, Great Wall, Loongson, Xingsen, Shennan, Honghe, Guanghe.

AI Applications: 1) Super Portals: Tencent, Minimax, Zhipu, Alibaba, iFlytek. 2) Transwarp, Decai, Meinian, Supcon, Zhuoyi, Kunlun, etc., for AI infrastructure, high growth, and high barriers.

Others: Aerospace era, Embodied AI, etc.

Risk Warnings Risks of intensifying industry competition; risks of technological iteration falling short of expectations; risks of cyclical fluctuations in downstream capital expenditures for specific industries.