GLM 4.6 Fuels Tenfold Rise in Zhipu AI's Overseas Paid Users as Coding Tools Go Global

100,000 paying customers outside China

Zhipu AI says its overseas paid user base has grown tenfold in two months, reaching about 100,000 monthly API customers alongside roughly 3 million free chatbot users outside China. The Beijing startup behind the Z.ai brand is turning demand for affordable coding assistance into a larger international business after launching GLM 4.6 in late September.

As of November 5, 2025, the company's expansion rests on a combination of price, practical programming performance and access through tools developers already use. GLM 4.6 is designed for coding, reasoning and agent workflows, in which an AI system carries out a sequence of tasks using software tools rather than simply answering a question.

Li Zixuan, Zhipu's head of global operations, said overseas demand is growing faster than demand at home. He described a strategy of investing across regions rather than giving priority to a single market.

"We want to grow in every market because growth overseas is exceeding growth domestically," said Li Zixuan, head of global operations at Zhipu AI.

The figures show two different routes into Zhipu's services. Paying API customers access the model through a software interface that allows applications to send requests and receive responses. Free chatbot users interact with the service directly. The 100,000 paying customers and 3 million free users therefore represent different types of usage within the company's overseas business.

GLM 4.6 quickly entered the daily work of developers through its relatively low price, straightforward integration and ability to handle real programming tasks. Individual programmers and small teams have been a central part of that appeal, especially when they need useful assistance without the expense of leading US systems.

How existing coding tools helped it spread

Claude Code, the coding tool backed by Anthropic, added access to GLM 4.6. Cline, Roo Code and Kilo Code also supported the model, giving developers several ways to use it without abandoning their existing workflows.

Kilo Code said GLM 4.6 produced the fastest adoption its platform had seen. Token usage rose 94 times in 12 days, with the model attracting users as a low cost option that was "good enough" compared with leading US systems.

A token is a small piece of text, often a word fragment, that a model processes when reading or writing. Kilo Code's figure measures the growth in text processed through the model. Zhipu's tenfold increase concerns paying users over two months. Together, the figures describe growth in both customer numbers and activity, using different measures and periods.

Support inside familiar tools makes trying a new model simpler. A programmer can switch providers within the same coding agent, compare how they handle a problem and choose according to speed, quality and cost. Those agents can manage tasks involving web search, repository operations and other tools across several steps.

The main dates and periods in Zhipu's expansion are:

• Late September 2025: Zhipu launched GLM 4.6 for coding, reasoning and agent workflows.
• The two months leading up to the November 5 account: Zhipu said its overseas paid user base increased tenfold.
• A period of 12 days during adoption: Kilo Code said token usage for GLM 4.6 rose 94 times.
• Recent weeks: new tools associated with Cognition AI and Cursor drew scrutiny over the origins of their underlying models.
• 2026: Zhipu is targeting a stock listing.
• 2030: chief executive Zhang Peng discussed this year as a possible marker for AI surpassing human ability in some areas.

What a larger context window changes

GLM 4.6 extends its context window to 200,000 tokens. The context window is the amount of text a model can consider at once, including instructions, documents and code. A larger window allows it to work with more files before it starts producing an answer.

For programming, that capacity can accommodate long chains of functions, configuration files and test suites. Developers working on larger repositories need the model to connect information across files, rather than repeatedly supplying small pieces of a project. More room for that information can reduce the exchanges needed to complete a task.

Developers report that GLM 4.6 works better with repositories containing multiple files and can plan longer sequences when building features. It makes fewer context mistakes on large projects because it can read more files at once. Careful prompts and a suitable strategy for using tools still remain part of the workflow.

Zhipu also describes improvements in coding quality, reasoning and tool use compared with GLM 4.5. The newer model fits more smoothly into agent frameworks and supports tool calling during inference, the process in which a trained model generates its response. That allows it to request software operations while working through a task.

Front end development, the part of software that users see and interact with, is another area of improvement. GLM 4.6 produces more polished pages and cleaner components than its predecessor. Developers say complete pages can require fewer edits, although complex styling and unusual cases still benefit from human review.

Performance and the cost of finishing a task

Zhipu's evaluations across agents, reasoning and coding show gains over GLM 4.5. On internal and public tests, GLM 4.6 is competitive with many widely used models, while the company acknowledges that it still trails Claude Sonnet 4.5 on some coding benchmarks.

In CC Bench agent tasks, GLM 4.6 achieved a 48.6 percent win rate in direct comparisons with Claude Sonnet 4, putting it close to parity in that evaluation. Zhipu expanded the tests to include more difficult work involving front end development, tool building, data analysis, testing and algorithm problems.

Human evaluators worked with the models in isolated environments to assess realistic workflows involving multiple exchanges. This approach looked at how systems completed a sequence of programming tasks, rather than judging only a single answer.

GLM 4.6 completed jobs using about 15 percent fewer tokens than GLM 4.5. Because API charges often depend on token consumption, using fewer tokens for the same task can reduce the bill. It can also make the interaction feel faster.

Developers still have to balance benchmark results against everyday reliability. A model that handles debugging, backend functions, integrations and interface components at an acceptable cost can be useful even when another model leads on particular tests. Backend functions are the parts of an application that process information and carry out work behind the visible interface.

The price pitch to developers

Zhipu markets its coding plan as providing performance at Claude's level for a fraction of the price. Its promotional comparison puts the cost at roughly one seventh of a comparable tier on a leading US service, with a larger usage quota for subscribers.

The exact value varies with usage and region. The company's pitch nevertheless addresses a practical concern for freelancers, students and startups: whether an AI assistant is affordable enough to use every day after an initial trial.

Developer reviews have described GLM 4.6 as cost efficient. Even when its output needs occasional corrections, its speed and price can make it a workable choice for routine programming. For teams with tight budgets, the relevant calculation includes both the subscription or API charge and the work needed to correct an answer.

Lower token consumption adds another part to that calculation. A cheaper rate and fewer tokens needed to finish a job can both contribute to lower costs. Zhipu's efficiency comparison with GLM 4.5 gives customers a way to consider the expense of completing work, rather than looking only at the advertised price.

Subscribers to Zhipu's coding plan are automatically moved to GLM 4.6. For applications, the configuration change amounts to selecting the new model name. That simple upgrade path helps existing users try the model without rebuilding their setup.

Open weights and control over deployment

Developers can access GLM 4.6 through the Z.ai API or routing services such as OpenRouter. They can also run its open weights locally. Weights are the numerical values learned during training that determine how a model processes information and produces responses.

Zhipu has published the weights on Hugging Face and ModelScope. It supports vLLM and SGLang, commonly used inference stacks that provide the software needed to run a model. These options appeal to engineering teams seeking greater control over deployment.

The company's GLM 4.6 blog describes the technical changes and recommended settings for coding and tasks assisted by tools. Zhipu also maintains a model page and documentation through its official channels.

Running the model locally, on company premises or in a chosen cloud region gives customers more say over where their code and documents are processed. Teams can also customize their setup instead of relying entirely on a remotely operated service.

Zhipu's openness, pricing and integration reinforce each other. Open weights provide deployment choices, familiar coding tools reduce the effort of adoption, and lower charges make repeated use easier to justify. The company is using that combination to compete for developers who want dependable assistance and control over their spending.

A debate over the origins of new coding models

Cognition AI's new model, SWE 1.5, and Cursor's Composer tool drew attention in recent weeks after users noticed behavior that appeared to point toward Chinese models. One tool generated reasoning logs in Chinese. Another acknowledged being built on a leading open model without identifying it.

Zhipu responded publicly to questions about model origins with a statement expressing its belief about SWE 1.5:

"We believe SWE 1.5 used our latest GLM 4.6 as the base model," Zhipu AI said in a statement responding to questions about model origins.

The base model behind the tools had not been publicly confirmed in the account available on November 5. Zhipu's statement sets out the company's position, while the unnamed foundation of the other tool remains part of the discussion.

The debate centers more on norms around disclosure and credit than on law. Open licenses can permit commercial reuse, including reuse without attribution. This allows models with open weights to become the foundation for products made by other companies.

Florian Brand, a researcher at Trier University who studies open models, described the role of the additional training and engineering that turn a base model into a commercial product:

"The fine tuning is the sauce," said Florian Brand, referring to the custom training and engineering that turn a base model into a polished product.

Fine tuning means adapting an existing model with further training for particular tasks or styles of work. Data pipelines and system engineering can also shape how a product performs. A commercial coding tool may therefore draw its value from the foundation model and from the work added around it.

Why Zhipu is looking abroad

Zhipu's overseas push reflects strong demand for affordable coding AI alongside pressures in its home market. In China, the company faces a slower pipeline of public sector business and increased competition. Management says expansion abroad is now faster than domestic growth.

The company, founded in 2019 as a spin out from Tsinghua University, has become one of China's most visible AI firms. It is valued at roughly 40 billion yuan and is preparing for a stock listing targeted for 2026.

Zhipu is concentrating on enterprise deals and a coding subscription aimed at developers to build more predictable revenue. It is not yet pursuing a mass market consumer subscription business to rival US platforms, although its free chatbot helps bring new users into the service.

Geopolitics affect purchasing decisions. Zhipu says customers in third countries are comfortable using its services when data can be hosted locally or when the company commits to zero data retention. Zero retention means the service does not keep customer data after processing under the agreed arrangement.

Local hosting and retention commitments are becoming standard ways for AI providers to address privacy concerns across borders. They are especially relevant for customers sending sensitive code and documents to a model.

Chinese AI firms also operate amid export controls, supply chain limits and regulatory scrutiny in multiple jurisdictions. These conditions affect training and deployment plans. Open weights give customers options for where and how a model runs, while buyers still assess support over time, service commitments and compliance features.

Competing in a crowded coding market

GitHub Copilot helped popularize AI coding assistance. Models from OpenAI and Anthropic power premium tools with strong reasoning and context management, while Chinese systems including DeepSeek and Qwen have advanced quickly.

Startups in many regions are also developing agents that work inside programming environments. These tools combine searching, analysis, code changes and tests in a continuing loop, making the underlying model part of a larger software workflow.

Zhipu's position in that market rests on three features: a relatively low price, access to open weights and support inside popular coding agents. Together, they offer a route to steady growth among developers who value reliability, deployment choices and a manageable bill.

The company is competing on the usefulness of the complete experience. A developer needs the model to understand the project, carry out a task with tools and produce code that can be reviewed and tested. GLM 4.6's larger context window, improved tool use and lower token consumption each address part of that process.

What the chief executive expects from AI

Zhang Peng, Zhipu's chief executive, has argued that discussions of artificial superintelligence often blur important details. He treats timelines as moving targets and focuses on whether models can outperform humans on particular jobs and turn those gains into real productivity.

At a recent launch event, Zhang described his view of the coming years:

"Achieving or exceeding human intelligence by 2030 might mean surpassing humans in certain areas, but still falling short in many others," said Zhang Peng, CEO of Zhipu AI.

Zhipu's current product strategy follows that practical focus. Enterprise services, the developer coding plan and an accessible free chatbot serve different users, while the immediate goal is to help people build features, fix bugs and work faster at a cost they can sustain.

What to Know

• Zhipu says overseas paid users increased tenfold in two months to about 100,000 monthly API customers, alongside roughly 3 million free chatbot users.
• GLM 4.6 launched in late September and became available through Claude Code, Cline, Roo Code and Kilo Code.
• Kilo Code said token usage rose 94 times in 12 days, its fastest adoption of a model.
• GLM 4.6 has a context window of 200,000 tokens and completes tasks using about 15 percent fewer tokens than GLM 4.5.
• Its 48.6 percent win rate against Claude Sonnet 4 on CC Bench put it near parity in those tasks; it still trails Claude Sonnet 4.5 on some coding benchmarks.
• Zhipu promotes a coding plan costing roughly one seventh of a comparable leading US tier, with value depending on usage and region.
• Open weights, API access and support for vLLM and SGLang give customers several deployment options.
• Zhipu believes SWE 1.5 used GLM 4.6 as its base model; the underlying model had not been publicly confirmed in the November 5 account.
• The company is valued at roughly 40 billion yuan and is targeting a stock listing in 2026 while investing in overseas enterprise and developer demand.