Coding is still catching up, while the battle for Work has already begun.

In October 2024, U.S. AI coding startup Bolt.new went live. Four weeks later, its annualized revenue surged to $4 million. Anthropic President Dario excitedly called Bolt's CEO, saying, 'You are our fastest-growing customer ever.' Bolt adopted Anthropic's newly launched model, Claude 3.5 Sonnet, pushing Anthropic's GPUs to full capacity.

Around the same time, an AI programming team at a major Chinese company was 'relatively idle, with no need for overtime.' Another enterprise took on custom projects, with multiple teams beginning on-site work. Several other tech giants were preparing for the Chinese New Year red packet rush, overshadowing the highly valuable AI coding sector. 'There are many contingent factors at play,' summarized a senior industry insider.

As of August 6th this year, on the developer platform Vercel, Anthropic accounted for 24.9% of token consumption and 71.8% of spending; its annualized revenue reached $47 billion in May, with Cursor exceeding $4 billion. In contrast, none of Alibaba, Tencent, or ByteDance have disclosed revenue from their coding products.

The gap is caused by multiple factors. Mid-2025 marks a dividing line. Before then, three forces simultaneously widened the gap in Vibe Coding between China and the United States—strategic bets, business models, and data flywheels. Chinese companies took nearly three years to catch on and accelerated their pace. From its inception, Anthropic bet on programming, which 'seemed like just a vertical niche market,' and it is now becoming the greatest value proposition in AI. Now, the battleground has expanded from coding to work, from 30 million programmers to 1 billion knowledge workers.

01 Sino-US Gap in Vibe Coding

'I must correct your view. It's not that domestic Coding is inferior; Claude is just too powerful. Domestic companies are also capable,' a senior industry insider told Shuzhi Qianxian. The domestic bottleneck lies in computing power. 'Anthropic has more GPU cards than all domestic vendors combined.' However, there is another voice in the domestic industry, arguing that the gap begins with foundational innovation.

Before mid-2025, ByteDance, Alibaba, Tencent, Zhipu, and Baidu primarily benchmarked against OpenAI. 'We chased general capabilities, text-to-image, text-to-video, C-end, B-end—whatever OpenAI did, we did,' said a Zhipu insider. At the time, OpenAI was valued at $300 billion. 'In comparison, programming seemed like just a vertical niche market.' Deepseek's first-generation model focused on coding but later shifted direction.

In contrast, Anthropic's oral history, 'The Making of Claude Code,' released in July 2026, revealed that the company considered Coding a strategic direction from its inception. Its first product in 2021 was a VS Code programming assistant. Notably, in early 2022, Anthropic's reinforcement learning team built a platform to train models capable of autonomously completing software tasks, believing that 'the path to AGI likely requires large-scale automated software engineering.' Between 2023 and 2024, an internal tool called 'clide' emerged—a 'crude' product that became the prototype for Claude Code, defining the shape of all coding CLIs today.

Huang Tiejun, Chairman of the Beijing Academy of Artificial Intelligence, told Shuzhi Qianxian that when Anthropic trained its models, code tokens accounted for 4.2 trillion, more than one-third of the total, with about half sourced from commercial software code. 'I believe all companies developing large language models initially valued code, but to varying degrees,' he said bluntly. 'What deserves reflection is that the digital world's impact on us is often underestimated. Modern society runs on the power grid, which supports an information network. Aren't many of our information systems computer code? OpenAI also regrets not paying enough attention to this area and being overtaken.'

Beyond strategic bets, the gap in business models is even more pronounced. As early as 2024, China's first batch of large model companies received signals that 'coding could generate revenue.' Baidu Intelligent Cloud insiders told Shuzhi Qianxian that in procurement records that year, AI coding applications ranked high in transaction value and volume, with the top four demand industries being finance, pan-tech internet, traditional software, and manufacturing. iFLYTEK Chairman Liu Qingfeng also mentioned that financial clients extensively used model-generated code.

However, a significant portion of these projects involved custom development and on-site services. This led Alibaba's Tongyi Lingma to develop standard products while also undertaking custom projects, resulting in high investment but limited returns. Zhipu expanded its government and enterprise project teams in 2024, dedicating significant manpower to projects. Customization was seen as arduous work with a clear growth ceiling. Without impressive growth curves, competing for resources within large companies became difficult. A Coding R&D insider at a major company recalled that at the time, their team of dozens was 'relatively idle.'

Overseas, however, growth curves were already soaring. 'Starting in mid-2024, the Coding sector exploded. Lovable, Boit.new, and GitHub Copilot all saw annualized revenues in the hundreds of millions of dollars, and not just one or two companies—over a dozen,' Chen Qiuwu, CTO of AI programming startup Kouding Technology, told Shuzhi Qianxian.

Click to add image description (max 60 characters)

These rapidly growing companies adopted subscription-based MaaS revenue models, where product capabilities were directly tied to model performance. Custom projects required additional work to enhance user experience. In contrast, Chinese tech giants did not fully shift to large-scale MaaS services until mid-to-late 2025 and 2026, transitioning from free customer acquisition to commercial charging.

Beyond business model differences, Anthropic had already begun closing the model, product, and data loop during this stage, which was the core reason for widening the model gap.

'Model training capacity and methods are no longer the bottleneck; it's just a matter of computing power and time,' a Huawei Cloud Coding product insider told Shuzhi Qianxian. 'The key is the data flywheel.'

Anthropic's Claude Code was the result of a series of cutting-edge explorations. In September 2024, after Boris Cherny, the soul of Claude Code, joined, colleagues rejected his handwritten code, suggesting he try the company's code tool, clide. Later, when Anthropic introduced tool use, Boris experimented by asking the model, 'What music am I listening to now?' The model autonomously wrote an AppleScript to query the media player, succeeding on its first attempt. This experiment convinced Boris that 'models should act as agents, with humans providing them with various tools to read, write, and run programs, rather than prematurely confining them to a fixed workflow.' This ultimately evolved into Claude Code. In the second half of 2024, a series of AI programming tools based on the Claude model quickly gained popularity, prompting Anthropic to rush out a product within two weeks.

In February 2025, Claude Code was released alongside the 3.7 Sonnet model. 'Anthropic signed agreements with users through Claude Code to collect data—developers' instructions, testing tasks, and code review focus areas. This data continuously trained its models,' explained a Huawei Cloud Coding insider. In reality, the GLM5.1 model released in 2026 had coding capabilities comparable to Claude Code. Now, the key differentiators are reasoning, scheduling tools, and project framework understanding—software engineering data that is simply unavailable on GitHub.

Even large companies' internal code repositories cannot fill this gap. 'The amount of code in large companies isn't particularly abundant, and the quality varies. Much of it is accumulated 'legacy code,'' said an industry insider. This is why, in late 2025, after Yao Shunyu joined Tencent, he proposed building a group-wide reinforcement learning infrastructure to feed real business data back to Hunyuan, aiming to establish a data flywheel.

Before mid-2025, China's Vibe Coding experienced a period of confusion and stagnation. Smaller players in the 'Hundred Models War' faced the dilemma that 'only large companies can afford foundational models' and scaled back model training to focus on applications. Even at large companies, those responsible for coding projects, while accurately identifying the direction, struggled to secure resources.

In January 2025, Ding Yu, head of Tongyi Lingma, hinted to Shuzhi Qianxian about the next phase's focus—autonomous programming, 'one-person companies,' and '20 programmers leading 10 AI programmers.' However, at the group level, the biggest AI battles in 2025 were over model open-sourcing and the 'universal assistant entry point,' with Doubao, Yuanbao, and Qianwen fiercely competing, overshadowing coding.

The real turning point came after mid-2025, when overseas players like Cursor and Claude Code reported soaring annualized revenues, prompting large companies' management to pay closer attention to coding. At Alibaba, Ding Yu reallocated personnel from multiple teams to compete directly with Cursor, developing Qoder in secret for several months before unveiling it at the August Cloud Town Conference.

02 Catching Up and Divergence

The turning point in 2025 came unexpectedly when DeepSeek launched R1, delivering a strong blow to the entire industry. Liu Jiang, Dean of Turing AI Institute and founding Vice Dean of Beijing Academy of Artificial Intelligence, recalled that before R1, the industry had grown somewhat complacent, but DeepSeek reignited hope by demonstrating the value of model training.

'Tang Jie sent an internal memo expressing regret that we didn't develop it but also renewed hope,' recalled a Zhipu insider. Tang later reflected that he had predicted large models would replace search, but instead, Google revolutionized its search with AI. After DeepSeek R1's emergence, 'this paradigm has largely reached its limit, leaving mostly engineering and technical challenges.' After many nights of debate, the team decided to focus on Coding and Agents.

Almost simultaneously, two kilometers away, Moonshot AI (Yuezhi'anmian) faced a choice. After losing to Doubao in a user acquisition war, Kimi halted large-scale product marketing and redirected resources to model development, planning to create China's first trillion-parameter model. Yang Zhilin said at the time, 'Startups must make their own bets to avoid protracted wars with competitors.'

However, the two companies took slightly different paths. A Zhipu insider recalled that the company was still aggressively pursuing government outsourcing projects for cash flow, as listing required it. In July 2025, Zhipu released GLM-4.5, integrating Coding, Agentic, and Reasoning capabilities into 'ARC,' a term that later became popular in the industry.

Zhipu also seized a window of opportunity: in September 2025, when Anthropic stopped serving Chinese users, it immediately launched a migration plan, offering Coding Plan at 20 RMB/month—one-seventh of Claude Pro's price. By December, media reported its ARR exceeded 100 million RMB, signaling strong MaaS momentum. During its March 2026 earnings call, CEO Zhang Peng explicitly positioned the company as 'China's Anthropic.'

Kimi focused more on the model layer. A month before releasing K3 this year, Yang Zhilin noted that programming scenarios accounted for over 90% of token consumption and that 'the underlying model still has many new variables.' He highlighted innovations in training the new model, such as replacing Adam optimizer (proposed in 2014) with MuonClip, linear attention (proposed in 2017) with a new architecture, and residual connections with attention residuals. 'Over the next 2-3 years, foundational technologies will be rewritten, with more innovative architectures emerging.'

Earlier, in February 2026, Anthropic accused Moonshot AI of distilling data through hundreds of fake accounts. However, multiple external researchers argued that K3's significant progress—with 2.8 trillion parameters—could not rely solely on distillation but on model scale, reinforcement learning, and engineering capabilities.

According to Stripe data, in January 2026, after Kimi released K2.5, its ARR surpassed $100 million, the first among the 'Six Little Tigers.' This revenue came from coding. In March, when Cursor released Composer 2, developers quickly discovered it was trained on Kimi K2.5's open-source model. Third-party industry data showed that after K3's launch, its API daily sales increased at least sixfold compared to before.

Large companies also acted during this period. At the August 2025 Cloud Town Conference, Qoder, Tongyi Lingma, and even a small team from TaoTian showcased vibe coding products. In September, Tencent released CodeBuddy. ByteDance launched Trae Enterprise Edition in late 2025, securing 'thousands of seats from major clients.' TRAE Personal Edition had the largest user base in China, with about 8 million users in June.

That was the closest large companies got to coding in 2025, but the sector remained 'narrow.' By autumn and winter, their attention shifted to another battle—competing for the AI dialogue assistant market, seen as the new traffic entry point in the AI era. DeepSeek, Tencent Yuanbao, ByteDance Doubao, and Alibaba Qianwen took turns vying for the top spot.

“Internally, integrating all Apps into QianWen is challenging, whether it’s coordination within the company or across different companies. Both us and ByteDance are doing the same thing—it’s a race to see who moves faster,” a senior Alibaba insider told Shuzhi Qianxian. QianWen subsequently integrated with Taobao, Alipay, Gaode, and Fliggy, while Doubao saw greater synergy with Douyin.

By the end of 2025, ByteDance secured the CCTV Spring Festival Gala partnership, Tencent prepared 1 billion yuan in cash red envelopes for distribution via Yuanbao, Alibaba invested 3 billion yuan to distribute milk tea through the QianWen App, and Baidu offered 500 million yuan in red envelopes to promote Wenxin Assistant. The competition for major platform entry points intensified, overshadowing attention on coding.

“In this round of the coding war, neither Chinese nor U.S. tech giants seem to be performing well,” observed an AI industry insider. Google lacks nothing in terms of models, computing power, talent, or code repositories, yet its coding products remain unremarkable. Flagship models from major companies must simultaneously address chat, search, multimodality, industry-specific needs... with programming being just one piece of the puzzle. Internal team competition fragments resources; unlike startups, they cannot go “all in” and must consider broader commercial positioning. To some extent, this is the result of multiple constraints.

It was once believed that only tech giants could afford to train foundational models. By 2026, Zhipu and Moonshot AI proved that startups could take risks, advance to the global frontier in model development, and continue pursuing AGI. Meanwhile, tech giants adopted multi-model strategies, swiftly transforming their proprietary models, Zhipu’s offerings, and Kimi’s open-source models into coding plans, cloud revenue streams, and various products. “Currently, among major companies, whoever holds the GPUs can quickly monetize coding,” a senior insider noted.

Overseas, companies like Cursor also leveraged Kimi’s open-source model weights, supplemented them with real-world feedback from their own Harness systems, and developed coding products—with the final achievement (results) attributed to Cursor.

In response, model startups are considering expanded monetization strategies. Yang Zhilin noted that more model companies are developing first-party Agent products. Take Anthropic’s Claude as an example: it does not Ultimate emphasis (overemphasize) long reasoning in every dimension but achieves strong product performance through Agent and Coding products. The industry observes that Kimi, on one hand, opens APIs and adapts to third-party coding tools while, on the other, launches Kimi CLI, Coding subscriptions, and first-party productivity products like Kimi Work.

By 2026, gaps persist, but confidence has shifted. “Catching up in coding isn’t difficult. Look at how quickly Kimi surpassed Zhipu—I believe QianWen and DeepSeek can do the same,” a senior Alibaba Cloud insider told Shuzhi Qianxian. “With better foundational models, our business scenarios are rich and diverse, and we have so many high-caliber engineers. Using real-world application data, we’ll quickly close the coding gap.” “We’re also training code models now—ByteDance-style, we’ll ramp up quickly,” a TRAE representative said.

“From a coding perspective, Chinese models lag behind Silicon Valley. If we rely solely on domestic models, they won’t match the capabilities of Claude Opus—that’s reality,” Liu Yi, Vice President of Tencent Cloud and head of CodeBuddy & WorkBuddy, admitted to Shuzhi Qianxian. However, he clarified that since mid-2025, “there’s no functional gap” in product capabilities. Their evaluations show that CodeBuddy’s Harness engineering scores “no lower than Openclaw” in SWE-bench tests.

“By focusing on product quality, user scenarios, and linking user and enterprise workflows, we can solve real problems step by step.” The battlefield has shifted: Chinese companies are now moving from coding to work.

03 The Battle from Coding to Work

In March, Tencent launched WorkBuddy in public beta. Three months later, Shuzhi Qianxian asked Wang Shengjie—the author of WorkBuddy’s first line of code and its first product manager—which product, CodeBuddy or WorkBuddy, holds greater value?

CodeBuddy is Tencent’s AI programming tool developed three years ago, while WorkBuddy targets knowledge workers who don’t write code. “It depends on the user group,” he replied. When asked if WorkBuddy would see wider adoption long-term, he answered, “Not necessarily—developers will still use CodeBuddy.”

A Tencent insider told Shuzhi Qianxian that coding products are more profitable at this stage, as many enterprises are still evaluating WorkBuddy’s token consumption costs, cost-saving potential, and opportunities for new business development.

Nevertheless, WorkBuddy has become Tencent’s most talked-about AI product this year. According to Analysys, by late June, it ranked first in China’s PC-end AI-native office agent market with 20.97 million monthly visits.

Behind this user data reversal lies a global industry shift within six months. In February 2025, Anthropic launched Claude Code; in January this year, it introduced Claude Cowork in research preview, targeting non-coding professionals.

Kimi’s Yang Zhilin calculated in June: programming scenarios account for over 90% of model token consumption, yet there are only 30 million programmers worldwide, compared to over 1 billion knowledge workers. “The next two to three years will see major paradigm shifts.” Pure coding capabilities will reach high proficiency within 12 months, but beyond that, agents must handle “end-to-end execution of many digital tasks.”

WorkBuddy is built on CodeBuddy’s foundation. Tencent Cloud’s Liu Yi explained that they spent over two years refining code products to create a stable, long-running “Coding Agent core,” which they then embedded into WorkBuddy. In contrast, many competitors’ work products, built on OpenClaw’s open-source projects, “fail to deliver results.”

Wang Shengjie recounted that in early January, he envisioned a Claude Cowork rival for Tencent’s internal workforce. The idea was to integrate Tencent Docs, email, Lexiang Knowledge Base, IMA, and even WeChat Reading and Music to bridge work-life silos. The foundation reused CodeBuddy’s compatibility with Anthropic’s ecosystem—plugins and Skills—allowing the platform to grow.

Tencent’s WorkBuddy momentum triggered a new round of competition among tech giants. From late July to early August, organizational reshuffles occurred almost simultaneously at three major companies. On July 20, Tencent’s QClaw team, which competed with Openclaw, was reassigned to Cloud Product Division VI, joining WorkBuddy under the same management. On July 30, ByteDance merged its Feishu product team with Doubao to form a new Doubao product team. On August 3, Alibaba launched QianWen Office in public beta, integrating QoderWork, MuleRun, and Wukong under DingTalk’s new CEO, Chen Yusen. Baidu launched DuMate this year, Kimi introduced Kimi Work, and 360 rolled out Nami Work.

“The industry now believes a single universal entry point will suffice, but product formats are still being explored. Look at Anthropic—it has Claude Code and CoWork, both on the same desktop,” Baidu’s DuMate chief architect, Li Jingqiu, analyzed for Shuzhi Qianxian. OpenAI follows a similar approach, with Work and Codex on the same desktop.

Chinese tech giants are also repositioning their products. Tencent offers both CodeBuddy and WorkBuddy; Baidu has Comate for programming and DuMate for office use. Alibaba’s Qoder, originally an all-in-one suite, now splits QoderWork from the suite, merging it with Wukong and MuleRun into QwenWork. Qoder (coding) remains a standalone product.

ByteDance maintains multiple channels: Doubao merged with Feishu to form the New Doubao product system, alongside the programming-focused Trae IDE and its derivative, Trae Work. An insider noted that merging Doubao with Feishu enhances appeal to white-collar workers.

“An organization has diverse roles—developers, product managers with technical knowledge, and functional departments. Some prefer zero-code solutions, others low-code, and some need high-code or vibe Coding,” Tencent’s Volcano Engine president, Tan Dai, told Shuzhi Qianxian in June. “The future may see more diversity or convergence—we’ll see as AI evolves.”

Why are all companies now shifting from coding to work?

A senior tech giant insider broke down the underlying logic: First, the user base is vast. With hundreds of millions of white-collar workers across diverse scenarios—administration, commerce, HR—“once adopted, there’s no turning back.” Second, white-collar loyalty far exceeds that of programmers. Programmers switch tools based on cost and performance; white-collar users, unconcerned with technical details, stay once they’ve built their workspace and found useful features. Third, and most critically—data. “Coding generates relatively little data; programmers prioritize privacy and use incognito modes.” Work scenarios differ: users upload 100 files for organization, creating vast, multidimensional data. In China’s walled-garden tech ecosystem, such data barriers hold long-term value.

“A programmer’s spending power doesn’t compare to a female white-collar worker’s,” he added. “This is the kind of battle tech giants love. Coding may not significantly impact market cap, but work products could amplify value by 100x or 1,000x.”

In reality, monetizing Work agents at the enterprise level remains uncharted territory. A ToB enterprise representative noted that work products enhance employee efficiency, differing from traditional enterprise procurement logic. Another ToB insider said models need refinement: “Improved meeting minutes in Feishu drive paid conversions. WorkBuddy’s knowledge base integration could lead to storage-based fees—domestic enterprise netdisks were once undervalued, but products like Feishu now monetize them effectively.”

Four years ago, Anthropic ignited the coding war, with competition centered on models, products, and data loops. Today, the Work battlefield is far larger, yet the core requirements remain the same—models, products, and data. All companies have entered the fray.

“We’re all acquiring users, encouraging rapid adoption, capturing preferences, refining skills, and retaining them,” a tech giant insider said.