9月22日，全球人工智能格局在几个小时内绕轴旋转。在北京，小米工程师进行了一场不同寻常的直播，直播了他们新的MiMo-V2.6型号的训练运行。在旧金山，Anthropic推出了其前沿Opus 5.5模型。大约一个小时后，OpenAI出人意料地发布了GPT-6 Sol和Luna型号，震惊了派对。
On September 22, the global artificial intelligence landscape spun on its axis within a matter of hours. In Beijing, Xiaomi engineers went live with an unusual broadcast, streaming the training run of their new MiMo-V2.6 model. In San Francisco, Anthropic unveiled its frontier Opus 5.5 model. About an hour later, OpenAI crashed the party with the surprise release of its GPT-6 Sol and Luna models.
对于试图跟上步伐的科技高管和开发人员来说，这只是人工智能战壕中又一个压倒性的一天。然而，当硅谷开始应对一些人所说的“模式疲劳”时，中国竞争激烈的科技行业正在经历这种现象。
For tech executives and developers trying to keep pace, it was just another overwhelming day in the AI trenches. But while Silicon Valley is beginning to grapple with what some have dubbed “model fatigue”, China’s hyper-competitive tech sector is experiencing the phenomenon on steroids.
这场为期一天的弹幕是在Z.ai、DeepSeek、腾讯控股、阿里巴巴集团控股和Moonshot AI的重大推出之后发生的，表明公司如何如此迅速地推出升级，以至于个别突破难以成为众人瞩目的焦点。
The one-day barrage – coming on the heels of major roll-outs from Z.ai, DeepSeek, Tencent Holdings, Alibaba Group Holding and Moonshot AI – shows how companies are churning out upgrades so rapidly that individual breakthroughs are struggling to command the spotlight.
科技分析师、Tech Buzz China时事通讯创始人马锐表示：“如果说有什么不同的话，那就是中国可能存在更多的模式疲劳。”“人们对这个速度感到疲惫，但他们也对生态系统的发展如此之快感到兴奋。”分析师表示，模型疲劳在全球两个最大的人工智能市场表现不同。
“If anything, there may be more model fatigue in China,” said Rui Ma, a tech analyst and founder of the Tech Buzz China newsletter. “People are exhausted by the pace, but they’re also excited that the ecosystem is moving this quickly.”Analysts said model fatigue manifested differently in the world’s two largest AI markets.
虽然美国企业正在努力解决增量性能增长是否证明转换模式的成本合理性，但中国的负担更重地落在了开发商身上。随着模型深入嵌入消费应用程序和云平台，快速的发布速度引发了残酷的价格战。
While US businesses wrestle with whether incremental performance gains justify the costs of switching models, the burden in China falls more heavily on developers. As models become deeply embedded in consumer apps and cloud platforms, the rapid pace of releases has triggered brutal price wars.
“在中国，疲劳是经济问题。每一代新车型都会重新设定价格下限，”人工智能研究公司The AI Capitol的创始人Lionel Sim说。
“In China, fatigue is about economics. Each new generation [of models] resets the price floor,” said Lionel Sim, founder of AI research firm The AI Capitol.
他补充说，客户越来越多地只对能力的真正飞跃或有意义的降价做出反应，而增量更新几乎没有引起注意。
Customers increasingly responded only to genuine leaps in capability or meaningful price cuts, with incremental updates drawing little attention, he added.
卡内基国际和平基金会非常驻学者Leia Wang表示，中国用户往往对价格更加敏感，这使得品牌忠诚度很难在模特层面建立。
Leia Wang, a non-resident scholar at the Carnegie Endowment for International Peace, said Chinese users tended to be more price-sensitive, making brand loyalty difficult to build at the model level.
王表示，模型层是“粘性最小的层”，并补充说，阿里巴巴等科技巨头试图通过将模型直接集成到更广泛的云生态系统中来锁定用户。这种经济压力在最近的发布周期中都很明显。
The model layer was the “least sticky layer”, Wang said, adding that tech giants like Alibaba attempted to lock in users by integrating models directly into broader cloud ecosystems. That economic pressure was evident throughout the latest release cycle.
DeepSeek在非高峰时段将其V4.1 Flash应用程序编程接口价格下调至每百万个缓存输入代币0.02元。Z.ai声称其GLM-5.3 Flash模型的表现优于上一版本，成本仅为10%，而阿里巴巴的Qwen 3.8-Flash每百万个输入代币收费1元（14美分）。
DeepSeek slashed its V4.1 Flash application programming interface prices to as low as 0.02 yuan per million cached input tokens during off-peak hours. Z.ai claimed its GLM-5.3 Flash model outperformed the previous iteration at 10 per cent the cost, while Alibaba’s Qwen3.8-Flash charges 1 yuan (14 US cents) per million input tokens.
小米在MiMo-V2.6上走了不同的路线，保持API价格稳定，同时推广“同价更高的智能”。它的直播还展示了对循环自我改进（RTI）的推动，这是人工智能系统在没有人类干预的情况下增强自身能力的过程。
Xiaomi took a different route with MiMo-V2.6, holding API prices steady while promoting “higher intelligence at the same price”. Its live stream also showcased a push towards recursive self-improvement (RSI), a process in which AI systems enhance their own capabilities without human intervention.
就连美国开发商也在适应价格紧缩。Anthropic将Opus 5.5的运营成本与其前身相比降低了40%，而OpenAI将其GPT-6 Sol和Luna型号的API价格相对于GPT-5.6的促销价格减半。
Even US developers are adjusting to the pricing squeeze. Anthropic cut operating costs for Opus 5.5 by 40 per cent compared to its predecessor, while OpenAI halved API prices for its GPT-6 Sol and Luna models relative to promotional rates for GPT-5.6.
现在每个版本都有太多的型号名称、基准声明和频繁的升级，无法获得持续的关注。然而，这并不意味着对人工智能的需求正在减弱赵波，《你好中国科技》尽管一些行业领导者越来越多地呼吁放慢速度，但无情的步伐仍在继续。就在推出Opus 5.5的前10天，Anthropic首席执行官达里奥·阿莫迪（Dario Amedei）敦促业界放慢前沿人工智能开发的步伐，理由是安全风险不断增加。
There are now too many model names, benchmark claims and frequent upgrades for every release to receive sustained attention. However, this does not mean demand for AI is weakeningPoe Zhao, Hello China Tech The relentless pace has continued despite growing calls from some industry leaders to slow down. Just 10 days before launching Opus 5.5, Anthropic CEO Dario Amodei urged the industry to temper the pace of frontier AI development, citing mounting safety risks.
这些呼吁在中国遭遇了强烈抵制，研究人员和官方媒体评论员将减缓发展的提议描述为美国试图维持其技术领先地位的战略手段。分析师表示，对中国开发者而言，放缓脚步根本不是一个选项。
Those calls have met strong resistance in China, where researchers and state media commentators have framed proposals to slow development as a strategic attempt to preserve the US’ technological lead. For Chinese developers, slowing down was simply not an option, analysts said.
“现在模型名称太多、基准测试宣称太多、每个版本发布都频繁升级，导致无法让每个发布都获得持续关注，”中国科技分析师、Hello China Tech通讯创始人Poe Zhao（赵博）表示。“但这并不意味着对AI的需求在减弱。”赵博指出，一旦AI系统投入生产环境，企业客户往往更看重稳定性、可预测的输出、集成成本、数据安全和技术支持，而非基准分数的微弱提升。只有当新模型能实现全新任务、大幅减少错误或大幅削减运营成本时，才具有商业意义。
“There are now too many model names, benchmark claims and frequent upgrades for every release to receive sustained attention,” said Poe Zhao, a China tech analyst and founder of newsletter Hello China Tech. “However, this does not mean demand for AI is weakening.”Zhao said enterprise clients tended to prioritise stability, predictable output, integration costs, data security and technical support over marginal improvements in benchmark scores once AI systems were deployed in production environments. A new model became commercially significant only when it enabled entirely new tasks, substantially reduced errors, or drastically slashed operating costs, Zhao added.
“中国是一个竞争极其激烈的市场，如果你停止发布产品，就有可能从讨论中消失，”Tech Buzz China的马睿（Rui Ma）说。为了在动荡中生存，中国科技公司正在开辟细分领域，而非仅在通用基准分数上竞争。
[China] is an ultra-competitive market, and if you stop shipping, you risk disappearing from the conversationRui Ma, Tech Buzz China To survive the churn, Chinese tech firms were carving out niches rather than competing solely on general benchmark scores, Tech Buzz China’s Ma said.
例如，字节跳动利用其短视频优势，迅速升级其Seedance视频生成模型。与此同时，阿里巴巴旗下的蚂蚁集团专注于医疗健康领域，其AI健康应用“AI问诊”（AQ）用户已达1.5亿，该公司表示。
ByteDance, for instance, has leveraged its short-video strengths by rapidly upgrading its Seedance video-generation models. Meanwhile, Ant Group, an affiliate of Alibaba, has focused on healthcare, with its AI health app AQ amassing 150 million users, the firm said.
随着AI商业化加速，赌注依然很高。阿里巴巴报告称，AI相关产品收入已连续第12个季度保持三位数增长，推动其云与AI业务在6月季度同比增长45%。
The stakes remain high as AI monetisation accelerates. Alibaba reported that revenue from AI-related products grew at a triple-digit rate for a 12th consecutive quarter, boosting its cloud and AI division by 45 per cent year on year in the June quarter.
Z.ai和MiniMax也公布了销售激增的消息，因为它们正转向付费企业级产品，尽管巨额的基础设施支出仍使该行业大部分处于亏损状态。阿里巴巴是《南华早报》的拥有者。文/李宗明
Z.ai and MiniMax have also posted surging sales as they shift towards paid enterprise offerings, although massive infrastructure spending continues to keep much of the sector unprofitable. Alibaba is the owner of the South China Morning Post.