By 2026, AI is no longer just a tool for “question-and-answer dialogues” but begins to truly “complete tasks.” Behind this shift lies a series of technological changes, from model scaling to intelligence ceilings, reasoning efficiency, long-term tasks, and agent systems.

These changes converge in the narratives of the “pioneers.”

Anthropic is expected to publicly release its prospectus in late September, while Zhipu has released its first interim results since going public.

Both the Chinese and U.S. large model companies showcase the rapid pace of AI commercialization: According to Bloomberg, Anthropic's annualized revenue has surged from approximately $20 billion at the beginning of this year to about $65 billion by the end of July, achieving its first operating profit in the second quarter. Zhipu also saw significant revenue growth (+399.7%) in the first half of the year and a substantial increase in the number of users on its MaaS platform (+144%) by the end of August, with both volume and price rising.

Yet, paradoxically, neither company focuses on commerce; instead, they dedicate significant space to explaining how models can integrate into longer, more complex, and continuously paid workflows.

Anthropic emphasizes across various promotional channels how Claude Fable 5.1 handles coding, knowledge work, and long-term problem-solving tasks, achieving similar or even superior performance to Fable 5 at a lower cost.

Zhipu's interim report even reads like a technical blog: the first half covers revenue structure, token usage, and average API pricing, while the latter half is filled with terms like scaling deep, long-term training, and fully self-training.

Behind this lies an industry-level judgment: The keyword for AI is shifting from “model capabilities” to “task delivery.” Only through continuous breakthroughs in intelligence and cost can the next generation of productivity infrastructure be built.

The Inflection Point Arrives: Coding Validates Scalable Paid Demand for AI

After three years of AI application exploration, the earliest successful paid scenario is not in companionship or search but in coding.

According to Menlo Ventures, enterprise generative AI spending is estimated to reach approximately $37 billion in 2025, a 3.2-fold increase year-on-year. Among this, AI coding spending will rise from about $550 million in 2024 to approximately $4 billion, accounting for 55% of departmental AI spending.

Coding alone supports half of enterprise AI budgets.

Thus, the revenue trajectory of Anthropic's Claude Code: approximately $100 million in annualized revenue within five months of launch, reaching about $2.5 billion in February 2026, $8 billion in May, and nearly $15 billion by August.

Similar changes are occurring at Zhipu.

Why does coding lead the way? Because it simultaneously meets three conditions: task results can be automatically verified, value can be directly priced based on saved engineer hours, and the work environment is inherently digital, allowing complete model takeover.

After coding, the same logic is replicating to scenarios like legal, finance, cybersecurity, and operations.

Zhipu summarizes this shift with a “capability ladder”: Chat → Coding → Co-work → Autonomous AI. With each step up, tasks become longer and unit economic value higher.

During the reporting period, Zhipu's focus lies between Coding and Co-work, with attempts already underway in cybersecurity, legal, and finance.

So, what exactly have Zhipu and Anthropic done to sustain accelerated commercialization?

It is the changes happening at the frontier of the AI industry: With large model capabilities as the foundation, the value of AI applications is evolving from being a tool for answering questions to infrastructure for completing tasks.

However, achieving this requires more than just models.

Value Curve vs. Cost Curve: One Rises, One Falls

The evolution from Chat to Autonomous AI essentially represents an increase in value density.

This is evident in Anthropic's case.

According to Bloomberg, Anthropic's revenue exceeded $11.5 billion in the second quarter, achieving an operating profit of approximately $559 million for the first time. About 80% of its revenue comes from the enterprise side, with over 1,000 enterprise clients spending more than $1 million annually, and 8 of the top 10 Fortune 500 companies using its products.

At a stage where the AI industry generally exchanges losses for growth, these figures signify that as models enter codebase-level migration, enterprise agents, and long-task workflows, customers are willing to pay prices comparable to human labor costs for “delivered results.”

In its 2026 Agent Coding Trends Report, Anthropic further outlines the next step: AI coding agents are evolving into “armies of agents,” with long-running agents capable of independently building complete systems, forming a “smart collaboration” model between AI and humans.

Anthropic validates the first curve of AI commercialization—task value: The longer the task chain, the more tokens consumed, and the clearer the basis for charging.

The industry significance of this curve is that it provides an valuation anchor for the entire AI sector: When a company proves that a model's ability to deliver work can be directly translated into revenue, all companies advancing along the same path will be revalued.

The similarities between Zhipu and Anthropic were already evident in their technical roadmaps before financial data.

Zhipu's GLM series completed six generations of iteration from 4.6 to 5.3 in about 11 months, with its intelligence index rising from 32 to 60.

In the first half of 2026, Zhipu's revenue structure underwent a fundamental shift: project delivery-dominated income was rewritten by cloud-based usage, with the revenue share of open platforms and API services billed per call rising from 26.3% at the end of 2025 to 86.5%.

This shift was not a deliberate design but a natural evolution driven by improved model capabilities. Each unlocked more complex task scenario expanded the pool of sustainably paid demand.

Zhipu refers to this more explicitly as the “intelligence ceiling.”

As the intelligence ceiling breaks through, both volume and price rise simultaneously. By the end of August, token usage on the MaaS platform had grown over 40-fold from the beginning of the year, with Coding Plan usage increasing over 23-fold. Meanwhile, the average API selling price rose about 101%, and Coding Plan subscription prices also increased. The price hikes reflect an upward shift in task structure, with higher-value tasks accounting for a larger share.

Prices fluctuate around value, but for large model companies, promoting AI applications and achieving technological inclusivity requires considering a key factor: cost. While the value curve rises, the cost curve must fall in tandem.

This year, both Anthropic and Zhipu are betting on a more critical future: how to deliver the same intelligence at scalable costs.

Agent tasks continuously invoke models, search, code execution, and external tools. While task value increases, token consumption multiplies, creating a driving force for continuous infrastructure efficiency upgrades.

To this end, Anthropic relies on Amazon, Google, Microsoft, and long-term agreements to lock in computing power supply, leveraging U.S. tech giants' infrastructure to access affordable computing resources.

Meanwhile, Zhipu goes all-in on infrastructure: using 100,000 domestically produced chips for scalable low-cost inference, reducing unit token inference costs by 80% from the beginning of the year, and tripling end-to-end service performance. Management highlighted a key metric, the “computing power multiplier”: for every yuan invested in computing power, the corresponding open platform and API revenue increased 14-fold compared to the first half of last year.

From large model intelligence ceilings to computing infrastructure efficiency, as both curves move simultaneously, the AGI commercial value flywheel of “intelligence ceiling × token consumption” accelerates.

AI Companies Enter the System Capability Competition Phase

Capability precedes business.

Around capability enhancement, three migrations are occurring in the AI industry.

The first is technological migration. Zhipu emphasizes scaling deep and fully self-training, while Anthropic introduces Fable and Mythos, shifting the industry from parameter scale to long-term tasks and continuous feedback.

The second is cost migration. Zhipu builds domestic inference clusters, while Anthropic locks in global computing power. Model companies simultaneously control the intelligence ceiling and unit intelligence costs.

The third is commercial migration. Zhipu accelerates its shift to MaaS, while Anthropic enters coding and enterprise agents, transitioning AI from providing answers to delivering work.

These three migrations point to a commonality: The competitiveness of AI companies is difficult to establish through a single generation of models alone. Models, users, tasks, computing power, and data must mutually supply each other.

In the past, the market primarily judged a company based on model leaderboard rankings and API call volumes. However, since 2026, the time window for single-point model advantages has increasingly narrowed, with leading positions never secure. Competitors can catch up within months. Companies that sustain leadership rely not on a single generation's brilliance but on a system that ensures each subsequent model outperforms the last.

Currently, Zhipu and Anthropic are among the few in the industry that integrate models, infrastructure, task environments, and commercialization into a holistic design.

Zhipu's system is built around three layers: at the computing layer, 100,000 domestically produced chips enable scalable inference, reducing unit token costs by 80% and increasing API revenue per yuan of computing investment by 14-fold; at the data layer, fully self-training allows models to evaluate and train each other, with GLM-5.3 achieving over 50% end-to-end completion rate improvement over 5.2 with the same architecture parameters, solely through expanded post-training scale; at the evolution layer, the next-generation GLM-6.0 is positioned as “self-evolving,” gradually detach ing human intervention and moving toward autonomous optimization.

Anthropic follows a similar logic: at the computing layer, it partners with SpaceX and Google to secure large-scale GPU supplies; at the data layer, it uses “Outcomes Loop” and “Dreaming” functions for AI self-assessment and background trial-and-error; at the evolution layer, it explicitly proposes “recursive self-improvement,” with AI already assisting in building the next-generation AI system—Claude 8 helping construct Claude 9, further accelerating the pace.

Essentially, both companies are doing the same thing: establishing a closed loop of “computing power → model → task → data,” making each iteration faster, cheaper, and stronger than the last. The future industry landscape will likely hinge on which flywheel spins faster, more steadily, and longer.

Additionally, there is a crucial factor: security.

As intelligence enhances, security and compliance capabilities are becoming foundational infrastructure for leading AI companies.

Anthropic emphasized in its August risk report that AI safety permeates the entire R&D and deployment lifecycle, while Zhipu opts for controlled openness of its most sensitive capabilities, with third-party independent evaluations before open-sourcing, vulnerabilities disclosed through national vulnerability databases, benchmark scores self-reported, and risk judgments left to external confirmation.

For B2B businesses, this “intelligence-cost-security” system is itself part of the productivity solution.

Conclusion

In 2026, the AI industry is nurturing the next generation of tech company business models.

Anthropic proves the validity of the task value curve with $65 billion in annualized revenue and the explosion of coding scenarios, while Zhipu demonstrates the validity of the scalable path with a 2736% API revenue growth rate and autonomous cost curve control.

Both companies write similar solutions on the same exam paper, jointly pointing to an upward spiral: “stronger model → more complex tasks → more tokens and revenue → more training and infrastructure investment → lower inference costs and more real feedback → next-generation model.”

Judging the value of AI companies now has new coordinates: whether model upgrades lead to task completion rate improvements, whether task capability improvements translate into call volumes and prices, whether call growth accompanies unit token cost reductions, and whether real tasks can be precipitate ed as training assets for the next-generation model.

Anthropic's IPO valuation could reach up to $2 trillion. If realized, it would rank among the largest global IPOs, potentially rewriting valuations for AI sectors in Hong Kong and U.S. stocks.

The whistle for the second half of AI commercialization has been blown.

Source: Hong Kong Stocks Research Society