转自：中国证券报
新华财经北京9月9日电（记者葛瑶）近日，OpenAI推出了全新一代人工智能（AI）模型GPT-6 Astra，显著增强了模型操控计算机及处理复杂任务的能力。英伟达CEO黄仁勋随即在社交媒体上宣称"通用人工智能（AGI）已然降临"，引发市场广泛关注。受新模型发布影响，A股算力硬件板块及韩国存储芯片龙头企业纷纷走强，AI基础设施需求再度成为焦点。
研究机构指出，模型能力的持续提升正不断拓宽AI的应用边界，企业及用户对使用成本和投入产出比的关注度日益提高。随着模型加速融入实际业务流程，算力需求、任务执行效率以及商业化落地进程已成为观察产业发展的核心指标。
模型执行能力提高
不久前，OpenAI正式推出GPT-6 Astra。根据公司披露，该新模型融合了预训练与强化学习等领域的最新研究成果，在计算机操作、软件工程及科学研究等方面实现了显著突破，能够跨多种软件和工具协同完成专业任务。
从官方演示来看，Astra具备填写在线表格、更新客户管理系统及安排日程等功能，还能分析科研数据、制作图表、构建网站并进行质量检测。用户只需设定目标，模型便能自主执行多步骤工作流程，产出文档、表格和演示文稿，其应用场景已从日常办公延伸至专业生产领域。
黄仁勋通过社交平台发文向OpenAI团队表达祝贺，并强调"AGI已经到来"。中国银河证券分析指出，Astra的关键突破主要体现在智能体执行能力与深度推理两大方面，模型能够自主规划任务步骤、灵活切换工具并纠正错误，大幅降低了对人工逐步指令的依赖程度。
模型在复杂任务执行能力方面的进步，进一步强化了市场对AI基础设施及存储芯片需求增长的预期，推动亚洲半导体板块热度攀升。9月7日，创业板指攀升3.41%，科创50指数上扬2.42%。存储芯片、光模块、印制电路板等细分领域表现抢眼，剑桥科技、光迅科技、景旺电子等强势涨停，源杰科技涨幅超过16%，长光华芯涨幅超10%。然而次日（9月8日），A股算力板块出现一定回调。
韩国市场半导体板块同样在近期呈现上涨态势。9月7日，三星电子与SK海力士分别录得5.68%和8.26%的涨幅。9月8日，SK海力士继续上涨0.56%。
据中证海外资讯报道，高盛亚太区首席股票策略师慕天辉（Timothy Moe）近期指出，全球范围内数据中心建设的激烈竞争将引发内存及NAND闪存芯片的严重供应紧缺，价格上涨趋势将在明年进一步加剧，从而持续推升半导体的市场价值。
成本与效率成为核心评判标准
研究机构认为，伴随模型能力的加速提升，AI的应用场景有望持续扩展，而完成实际任务所需的成本与效率已成为评估模型商业价值的重要标尺。
开源证券指出，尽管Astra的API定价较上一代模型有所上调，但从部分测试结果及客户实践来看，该模型在降低词元消耗、减少人工修正需求以及缩短任务处理周期等方面展现出明显效率优势，部分任务的整体成本出现下降。该机构认为，当成本评估方式从Token单价转向任务完成综合成本后，企业级智能体部署的经济可行性有望得到进一步提升。
瑞银证券中国互联网行业分析师熊玮认为，在审视中国AI及大模型产业时，应将重心放在模型能力、Token投入产出比以及模型厂商的商业化推进情况上。
就模型能力而言，熊玮指出，编程能力不仅是已获验证、具备清晰能力提升路径和商业化潜力的方向，还能带动工具调用、长上下文等更广泛能力的增强，使模型能够替人类完成更困难、更复杂且更具价值的任务。相应的市场空间也正从编码和软件开发领域，逐步向更广泛的知识工作者生产力场景拓展。
与此同时，企业对AI应用的态度正发生转变。熊玮表示，今年初不少企业鼓励内部扩大模型使用规模，但运营一段时间后发现相关费用可能超出预算，且Token消耗所产生的经济价值难以准确评估。自今年年中起，企业、开发者及用户开始更加注重成本优化，在绝对性能表现与性价比之间寻求平衡点。
熊玮认为，全球大模型市场或将进一步呈现分层格局。对于价值高、敏感度高及质量要求严苛的任务，可能仍需依赖全球顶尖模型；而企业日常运营中大量低风险、高重复性的任务，则可考虑采用成本更低的版本。中国开源模型能力持续增强，叠加价格优势，正逐渐成为企业构建工作流程和执行任务时的可选方案。
在商业化方面，熊玮指出，开源模型虽然变现路径不及闭源模型直接，但依然存在较大的发展空间。模型厂商既可通过官方API提供服务，也能在大规模企业部署过程中，通过收入分成、项目服务等方式实现营收。
熊玮表示，目前全球AI应用与商业化变现仍处于相对初期的阶段。随着模型能力持续提升，中国模型在训练和推理端的成本优势有望进一步凸显，未来商业化拓展空间值得持续关注。
编辑：张瑶
声明：新华财经（中国金融信息网）为新华社承建的国家金融信息平台。任何情况下，本平台所发布的信息均不构成投资建议。如有难题，请联系客服：400-6123115本文来源：https://finance.sina.com.cn/money/bond/2026-09-09/doc-inirense5765302.shtml
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Source: China Securities Journal
Xinhua Finance, Beijing, September 9th — Recently, OpenAI launched the new generation artificial intelligence (AI) model GPT-6 Astra, significantly enhancing the model's ability to control computers and handle complex tasks. NVIDIA CEO Jensen Huang immediately declared on social media that "general artificial intelligence (AGI) has arrived," sparking widespread market attention. Following the release of the new model, the A-share computing hardware sector and Korean memory chip leaders saw gains, with AI infrastructure demand once again becoming the focus.
Research institutions point out that the continuous improvement of model capabilities is constantly expanding the boundaries of AI applications, with businesses and users increasingly focused on usage costs and input-output ratios. As models are more rapidly integrated into actual business processes, computing power requirements, task execution efficiency, and commercialization progress have become core indicators for observing industry development.
Model execution ability improves
Not long ago, OpenAI officially released GPT-6 Astra. According to the company's disclosure, this new model integrates the latest research results in pre-training and reinforcement learning, achieving significant breakthroughs in computer operations, software engineering, and scientific research, and can collaborate across multiple software and tools to complete professional tasks.
Based on the official demo, Astra has the ability to fill in online forms, update customer management systems, and schedule calendars, as well as analyze scientific data, create charts, build websites, and perform quality inspections. Users only need to set the goal, and the model will autonomously execute multi-step workflows, producing documents, tables, and presentations, expanding its application scope from daily office work to professional production fields.
Jensen Huang posted on social media to congratulate the OpenAI team and emphasized that "AGI has arrived." China Galaxy Securities analyzed that Astra's key breakthroughs lie in the execution capabilities of the agent and deep reasoning, as the model can autonomously plan task steps, switch tools flexibly, and correct errors, significantly reducing dependence on manual step-by-step instructions.
The model's progress in complex task execution has further reinforced market expectations for growth in AI infrastructure and memory chips, driving up the heat in the Asian semiconductor sector. On September 7th, the ChiNext index climbed by 3.41%, and the SSE STAR 50 index rose by 2.42%. Sub-sectors such as memory chips, optical modules, and printed circuit boards performed well, with Cambridge Technology, Accelink, and Jingwang Electronics among others hitting the daily limit up, while Source Optronics gained over 16% and Unikorn Electronics gained over 10%. However, on the next day (September 8th), the A-share computing power sector experienced a certain degree of pullback.
The semiconductor sector in the South Korean market also showed a rising trend recently. On September 7th, Samsung Electronics and SK Hynix rose by 5.68% and 8.26% respectively. On September 8th, SK Hynix continued to rise by 0.56%.
According to a report by China Securities Overseas News, Timothy Moe, Chief Equity Strategist for Asia Pacific at Goldman Sachs, recently pointed out that the fierce global competition in data center construction will lead to a severe shortage of memory and NAND flash chips, with price increases expected to intensify further next year, continuously driving up the market value of semiconductors.
Cost and efficiency become the core evaluation criteria
Research institutions believe that as model capabilities accelerate, AI application scenarios are expected to continue expanding, with the cost and efficiency required to complete actual tasks becoming an important yardstick for evaluating the commercial value of models.
Oppenheimer Securities pointed out that although the API pricing for Astra has been increased compared to the previous generation, from some test results and customer practices, the model shows significant efficiency advantages in reducing token consumption, minimizing the need for manual corrections, and shortening task processing times, with the overall cost of some tasks decreasing. The institution believes that as the cost evaluation method shifts from token unit price to the comprehensive cost of task completion, the economic viability of enterprise-level agent deployment is expected to improve further.
Xiong Wei, an internet industry analyst at UBS Securities China, believes that when examining the Chinese AI and large model industry, the focus should be placed on model capability, the input-output ratio of tokens, and the commercialization progress of model manufacturers.
Regarding model capability, Xiong Wei notes that programming ability is not only a verified direction with a clear path for capability improvement and commercialization potential but can also drive the enhancement of broader capabilities such as tool invocation and long-context understanding, allowing models to complete more difficult, complex, and valuable tasks for humans. The corresponding market space is gradually expanding from coding and software development to broader knowledge worker productivity scenarios.
At the same time, enterprises' attitudes toward AI applications are changing. Xiong Wei states that while many companies encouraged increasing model usage internally earlier this year, after running for a period, they discovered that the related expenses might exceed budget, and the economic value generated by token consumption is difficult to evaluate accurately. Starting from the middle of this year, companies, developers, and users have begun paying more attention to cost optimization, seeking a balance between absolute performance and cost-effectiveness.
Xiong Wei believes that the global large model market is likely to further present a tiered structure. For high-value, sensitive, and quality-demanding tasks, reliance on the world's top models may still be necessary; for the large number of low-risk, high-repetition tasks in daily enterprise operations, lower-cost versions can be considered. The continuous strength of Chinese open-source models, combined with price advantages, is gradually becoming an optional solution for enterprises building workflows and executing tasks.
In terms of commercialization, Xiong Wei points out that although open-source models have a less direct monetization path than closed-source models, there is still significant room for development. Model manufacturers can provide services through official APIs and achieve revenue through revenue sharing and project services during large-scale enterprise deployments.
Xiong Wei stated that the global AI application and commercialization are still in a relatively early stage. With the continuous improvement of model capabilities, the cost advantage of Chinese models in both training and inference is expected to become more pronounced, making future commercialization expansion worth sustained attention.
Editor: Zhang Yao
Source: https://finance.sina.com.cn/money/bond/2026-09-09/doc-inirense5765302.shtml If there are any copyright issues, please contact us at: 13951431913#139.com or submit feedback on the right.