Author: Daoge

Source: Zhibaidao

AI Coding is undeniably becoming the first commercialized sector within the AI landscape to achieve widespread adoption.

According to Market Research Future, the AI programming tools market is projected to surge from $15.11 billion in 2025 to $99.1 billion by 2034, with a CAGR of 23.24%.

At Meta's LlamaCon event in May, Microsoft CEO Nadella revealed that 30% of Microsoft's current code is AI-generated. Microsoft CTO Kevin Scott previously predicted that 95% of code would be AI-generated by 2030.

In China, Ministry of Industry and Information Technology data shows software business revenue reached RMB 9.8281 trillion in the first three quarters of 2024, up 10.8% YoY—all potential markets for AI programming adoption.

Faced with trillion-dollar commercial prospects, Chinese models are engaged in a fierce parameter pursuit race.

Take DeepSeek-V3.2 (launched December 1) as an example: its SWE Verified score for large model code engineering task resolution reached 73.1%, nearing Anthropic's Claude-4.5-Sonnet (launched September 29) at 74.9%.

However, focusing solely on numerical gaps may obscure the true battleground. Zhibaidao argues the decisive factor in Sino-US AI programming competition is shifting from parameter performance to ecological standards.

01 Chinese Models Embrace Open Source for Infrastructure; US Leans on Closed-Source Performance

Programming has long been defined as a rigorous, logic-driven discipline that translates human intent into machine-executable language. This precise causal logic makes it the ideal PMF (Product-Market Fit) domain for mainstream LLMs.

At AI Ascent 2025, Sequoia Capital declared AI Coding the first market to be disrupted, serving as a critical harbinger for AI adoption across other industries.

In this high-willingness-to-pay sector, China and the US have taken divergent paths:

The US pursues an elite closed-source route, leveraging superior model performance to attract capital markets and sustain staggering valuations.

AI programming tool Cursor recently secured $2.3 billion in Series D funding, with shareholders including Google and NVIDIA. Its parent company Anysphere's valuation tripled in four months to $29.3 billion. Meanwhile, B-side market dominator Anthropic's valuation soared to $350 billion.

Its newly launched Claude Opus 4.5 on November 25 achieved 80.9% in SWE-Bench Verified testing, surpassing Gemini-3 Pro and GPT-5.1 while becoming the first model to exceed 80 points—indicating AI's code correction capability now matches or exceeds human expert levels.

More damagingly, it slashed prices: Claude Opus 4.5's API pricing dropped to $5 (input)/$25 (output) per million tokens, cutting two-thirds from the previous generation.

In contrast, Chinese models have adopted an open-source + cost-effective strategy.

This year, DeepSeek's R1 breakthrough raised the benchmark for domestic models while establishing China's open-source foundation. After a year of competition, domestic model leaders Kimi, Qwen, and GLM now rival overseas top-tier models in parameter performance.

DeepSeek's newly launched V3.2 on December 1 achieved 73.1% in SWE Verified, nearing Anthropic's Claude-4.5-Sonnet (74.9%) and demonstrating the viability of open-source + cost-effectiveness.

These divergent strategies have created distinct user bases.

Fortune previously reported that many European and American executives favor the performance advantages of proprietary models from OpenAI, Anthropic, or Google.

The Asian market, however, prioritizes data sovereignty and cost control.

Yuan Jinhui, co-founder and CEO of Chinese AI cloud hosting provider SiliconFlow, stated his company developed technologies to run open-source models more economically, making task completion costs far lower than proprietary AI models. He added that fine-tuning open-source models with proprietary data for specific applications can outperform proprietary models while fully avoiding sensitive data or trade secret leaks.

Chen Yibang from Vertex Venture Holdings emphasized that while proprietary model vendors offer fine-tuning services based on customer data—often promising not to use it for broader training—the reality remains unverifiable.

Open-source models allow free downloading, modification, and integration, enabling startups to develop products more easily and researchers to improve models faster. Their widespread adoption is profoundly shaping AI's future trajectory—a logic now taking effect globally.

Singapore's National AI Programme (AISG) made a symbolic strategic shift by announcing its latest Southeast Asian language model "Sea-Lion" would abandon Meta's framework in favor of Alibaba's Qwen architecture. This marks China's open-source models beginning to break through globally via pragmatism.

Over the past year, Chinese teams' self-developed open-source AI models have increased their download share to 17.1%, surpassing the US's 15.8% for the first time. MIT and Hugging Face data show DeepSeek and Alibaba's Qwen models dominate Chinese model downloads.

02 China Builds Cars While the US Constructs Roads

As large model iteration slows, domestic models may eventually match or surpass US leaders in parameters. However, overseas advantages may extend beyond engineering capabilities to ecological moats.

Zhibaidao observes that top model vendors like Anthropic are attempting to define the HTTP of the agent era.

In February, Anthropic launched Claude Code—not merely a tool but a native IDE system. Unlike "shell" IDEs like Cursor, Claude Code achieves deep decoupling and reconstruction between models and development environments. It directly understands codebases, manages context, and invokes third-party tools.

Within four months, Claude Code attracted 115,000 developers. Menlo Ventures predicts it will generate $130 million in revenue for Anthropic.

Additionally, Google introduced the A2A open protocol to support model-to-model interoperability, addressing black-box states between opaque agent systems.

As each model excels in different areas, developers often need to combine multiple models for real-world applications. The A2A protocol enables users to invoke different large models to generate distinct agents for collaborative goal achievement.

While A2A facilitates cross-model collaboration, its complement is MCP.

In November 2024, Anthropic open-sourced MCP, defining standards for "how models connect to tools and data sources." Unlike Claude Code's simple linking, MCP enables models to autonomously decide which tools to invoke for complex tasks.

To address MCP's redundancy issues, Anthropic introduced Skills. Rather than creating new functions, Skills equip models with memory and processes—packaging business workflows, templates, and internal knowledge into modular Skill units for Claude to automatically invoke when appropriate. For developers, this adds a lightweight workflow layer to LLMs, improving controllability, flexibility, token efficiency, time savings, accuracy, and collaborative development. By integrating tools with models via Skills, developers expand model capabilities.

When agents can collaborate across platforms, players with the richest toolchains and operating systems gain legislative power. Ecological niche competition thus precedes technical route divergences.

Overseas large model vendors' mature understanding of B-side applications stems from America's more developed SaaS ecosystem. The US SaaS industry, originating in the 1980s, has established standardized workflows. Its heavy reliance on standardized APIs and plugin systems creates stronger demand for agent automation in cross-platform collaboration.

Chinese SaaS development lags by nearly a decade, with many enterprises operating non-highly-structured businesses—hindering domestic model vendors' promotion of ecological and standardized tools.

Fortunately, awakening has begun.

In August, Alibaba launched Qwen Code to catch up ecologically, revealing promising prospects for domestic models.

Qwen Code prioritizes developer experience and plans to expand IDE plugins, enhance tool invocation capabilities, and continuously improve engineering efficiency through functional accumulation.

Qwen Code is entering the core battlefield of "AI engineering" by attempting to develop workflow takeovers and establish its own rules.

From a long-term perspective, future success hinges not on single-model performance but on which country secures B-side standard-setting rights. While parameter improvements are rapid and cost-effective to chase, ecological maturity requires years of developer accumulation, interface standardization, and vertical industry understanding—factors that cannot be rushed.

The AI application generation tool landscape, like foundational model competition, is not a winner-takes-all market. Differentiation and coexistence will emerge, with domestic vendors aggressively pursuing both cost-effectiveness and ecological development. Only by establishing proprietary ecosystems and standards can Chinese AI truly surpass invisible moats.

*Featured image generated by AI