Terminal-Bench 4.0: GLM-5.3 Climbs to Third Place, Surpassing GPT-5.6 Sol

• Z.ai launched GLM-5.3-Flash, a low-cost AI model running entirely on domestic Chinese semiconductors, causing its stock to surge 8% to 12%.
• The model ranks 10th on the Artificial Analysis Intelligence Index but leads in agentic coding behavior, challenging the dominance of high-intelligence models.
• Anthropic's Claude model reports a 1,000% revenue surge, narrowing the competitive gap with OpenAI despite OpenAI's cost advantages.
• Market sentiment via Polymarket assigns Anthropic a 69% probability of leading the AI model race by 2026, reflecting a preference for capability over price .

Shares of Z.ai, a Beijing-based artificial intelligence firm, climbed between 8% and 12% in Hong Kong trading following the launch of GLM-5.3-Flash . The company stated that the model, previously previewed under the code name "Ox Alpha," operates entirely on homegrown semiconductors . Utilizing approximately 100,000 China-made chips, the system handles all online inference requests without reliance on foreign technology . This strategic move highlights progress in tech self-sufficiency amid ongoing U.S. export restrictions on advanced chip sales to Beijing .

The model features a 1 million-token context window and is natively multimodal, supporting image and video inputs . It ranks 10th on the Artificial Analysis Intelligence Index v4.1.1, outperforming competitors like DeepSeek V4 Pro Max in specific metrics . The model achieved a highly competitive cost of $0.045 per task, which Z.ai claims is roughly one-tenth the cost of Nvidia-based alternatives . It quickly secured the top spot for usage on the global OpenRouter platform in the week following its August 20 release .

How Does GLM-5.3-Flash Challenge the Intelligence Hierarchy?

GLM-5.3-Flash represents a strategic shift in AI economics, prioritizing agentic behavior and cost-efficiency over raw reasoning power. The model utilizes a sparse architecture with 320 billion total parameters but activates only 18 billion per token . This design cuts compute requirements while supporting a massive context window, enabling high throughput and low latency. Users report that it performs comparably to frontier models like Opus 4.8 in code generation and debugging .

The model excels in "behavior," defined as the ability to stay on task, handle mid-task steering, and self-correct during agentic workflows . While it may lag in pure reasoning benchmarks against smarter models, its efficiency in coding tasks makes it highly effective for high-volume applications . A critical driver of its low price point is its infrastructure, running exclusively on Huawei Ascend 910 BC chips . This reliance on Chinese hardware bypasses U.S. export controls, granting Z.ai access to massive domestic compute reserves .

Why Is Anthropic's Revenue Surge Reshaping Market Sentiment?

Anthropic is experiencing a massive revenue expansion, with reported surges of 1,000%, indicating strong market adoption of its Claude model . This growth suggests Anthropic is successfully gaining market share from competitors like OpenAI . Despite OpenAI's GPT-5.6 Sol currently ranking just one point behind Anthropic's Claude Fable 5 on the Artificial Analysis Intelligence Index, OpenAI holds a distinct advantage in cost efficiency . Sol is estimated to cost less than half as much per million tokens compared to Anthropic's offerings .

However, trading data from Polymarket indicates that traders currently assign a 69% chance of Claude being the top AI model by the end of 2026 . This disparity highlights a market preference for Anthropic's model capabilities or strategic positioning, even in the face of OpenAI's pricing power . The data suggests that while OpenAI remains a formidable contender, Anthropic's revenue momentum and perceived model superiority are key drivers in the ongoing AI infrastructure competition .

What Are the Investment Implications of Domestic AI Infrastructure?

The launch underscores China's strategic push for technological self-sufficiency in response to Washington's restrictions on advanced chip sales . While Nvidia has faced challenges selling to the Chinese market, domestic players like Huawei are ramping up production of alternative AI accelerators . Z.ai declined to identify the specific chip manufacturers used, and CNBC could not independently verify the claim of 100% domestic hardware utilization . Nevertheless, the move signals a shift toward cost-efficient inference infrastructure that bypasses reliance on foreign technology .

Rival MiniMax's shares climbed by around 3% in Hong Kong trading after reporting a 283% revenue surge in the first half of the year. This broader trend reflects Chinese AI developers investing heavily in domestic computing infrastructure to reduce dependence on U.S. semiconductors . The open-weight nature of GLM-5.3-Flash, available under an MIT license, further enhances its appeal for companies seeking hosting flexibility. This democratization of high-quality coding AI could lower barriers to entry for smaller developers and accelerate the adoption of autonomous coding agents .

Counterpoint Senior Research Analyst Ivan Lam noted that this launch reflects a broader trend of Chinese AI developers investing heavily in domestic computing infrastructure . The model is available under an MIT license, allowing unrestricted commercial use and further optimization . The broader implication is that the AI market is bifurcating, with high-cost, high-intelligence models remaining necessary for complex reasoning. However, cheap, high-behavior models like GLM-5.3-Flash will likely dominate high-volume, routine agentic tasks .