China's AI is bypassing giant LLMs and heading straight to the agent era

Background: a shift users can feel

Chinese firms and global labs alike are quietly changing how they deploy large language models — and users are noticing. In recent weeks, complaints surfaced in communities around Anthropic's Claude that the model feels "less smart": faster replies, shorter chains of reasoning, and a tendency to stop work early. It has been reported that a GitHub analysis of Claude Code logs found a roughly 67% drop in measured reasoning depth since a February update, a figure the author called an estimate rather than a precise measurement. The phenomenon even has a nickname in forums: "AI shrinkflation" — same name, less capability.

Anthropic has framed the change as engineering, not a capability loss. The company’s team reportedly says system-level adjustments — from tool-integration and prompt management to a new "adaptive thinking" mechanism that dials reasoning effort to task complexity — account for the difference. Is this optimization for efficiency, or a deliberate reweighting of resources to favor newer models? Reportedly, Anthropic also limited early access to its higher-capability Mythos preview to a handful of institutions, calling it a "generational leap" for critical-code and security tasks.

Why this matters for China and beyond

The technical trend lines matter particularly in China because local players are racing to ship practical, agent-style products rather than chase ever-larger monolithic models. Companies such as Baidu (百度), Alibaba (阿里巴巴) and Tencent (腾讯) are investing heavily in agent frameworks and lightweight orchestration that can run effectively on constrained hardware. Analysts note — and it has been reported that — Western export controls on cutting-edge AI chips and the rising cost of large-model inference have increased the appeal of agent architectures that split work across tools and smaller, specialized models.

The commercial and geopolitical consequences are immediate. Prioritizing new, high-value models for select partners while throttling older public endpoints can make capability gaps more visible and feeds suspicion among users. Regulators and customers will soon ask tough questions: are we trading transparency for efficiency? And will widespread adoption of agent systems concentrate power in the hands of those who control the orchestration stacks and the scarce compute resources? The next wave of AI in China looks less like a single giant brain and more like fleets of specialized agents — faster, leaner, and harder to audit.