For years, the global artificial intelligence race appeared to have a clear leader. American companies built the most capable models, controlled the largest computing infrastructure and attracted the bulk of investor attention. China was often viewed as a fast follower rather than a genuine challenger.
That perception is becoming increasingly difficult to sustain.
The latest evidence arrived with the release of GLM 5.2, an open-source model from Chinese AI company z.AI. Over the weekend, the model sparked widespread discussion across Silicon Valley after developers and researchers highlighted its ability to handle complex coding tasks and agentic workflows, capabilities typically associated with the most advanced systems from Anthropic and OpenAI.
The excitement surrounding GLM 5.2 is not simply about one model. It reflects a broader shift in the global AI landscape, where China is steadily reducing the performance gap that once separated its developers from their American counterparts.
What is China’s GLM 5.2 model?
Much of the recent attention around China’s AI progress has centred on GLM 5.2, the latest model from Beijing-based startup Zhipu AI. Rather than representing a complete redesign, the model builds on the company’s earlier work, with a particular focus on handling extremely large amounts of information while maintaining strong performance on coding and agentic tasks.
One of GLM 5.2’s defining features is its enormous context window, allowing it to process and retain far longer inputs than traditional AI models. This capability enables users to work with entire software repositories, lengthy research documents or complex multi-step workflows within a single session, reducing the need to repeatedly reload information.
Like DeepSeek’s models, GLM 5.2 has been released as an open-weight system. This means developers can download the model’s weights, run it on their own infrastructure and modify it to suit specific requirements. That approach differs sharply from the closed models offered by leading US firms such as OpenAI and Anthropic.
Quick Reads
US intelligence security agency loses access to Anthropic AI amid government dispute
US presses Meta to agree to AI reviews as security concerns rise
The model is built on a 753-billion-parameter architecture and is released under the permissive MIT open-source licence. Zhipu says it incorporates several efficiency improvements, including a new optimisation technique known as “IndexShare”, which significantly reduces the computing resources required to process long documents. It also features an upgraded speculative decoding system designed to accelerate response generation and offers different reasoning modes that allow users to balance speed and depth of analysis.
Industry reaction has been notably enthusiastic. Several prominent technology executives and developers have praised its coding capabilities, with some describing it as the first open model capable of serving as a reliable daily-use assistant.
Benchmark results have further fuelled interest. GLM 5.2 outperforms many leading open-weight rivals, including DeepSeek V4, and achieves results comparable to, and in some cases better than, frontier models from OpenAI and Anthropic. It has demonstrated particular strength in software engineering, tool use and long-horizon task completion, placing it among the most capable open AI models currently available.
The DeepSeek effect is turning into a trend
The conversation around China’s AI progress began to intensify after DeepSeek’s R1 model shocked the industry last year. The release challenged the assumption that cutting-edge AI development remained largely confined to a handful of US technology firms.
Since then, Chinese companies have continued to release increasingly capable models at a rapid pace. Rather than relying on a single breakthrough, the country’s AI ecosystem has demonstrated a pattern of consistent improvement across coding, reasoning and autonomous task execution.
GLM 5.2 is the latest example. Its emergence has reinforced a growing belief among industry observers that Chinese developers are now competing much closer to the frontier than many policymakers and investors previously assumed.
Recent research appears to support that view. Stanford University’s AI Index Report found that Chinese models significantly improved their quality over the past year, narrowing performance differences with leading US systems and, in some cases, approaching parity on widely used benchmarks.
The debate is no longer centred on whether China is catching up. Increasingly, the question is how quickly.
Why experts disagree on the size of America’s lead
Despite the growing momentum behind Chinese AI, there remains little consensus on how much of an advantage the United States still holds.
David Sacks, who previously served as the White House’s AI and cryptocurrency adviser, recently argued that the US lead may be as little as six to nine months. Such estimates have fuelled concerns among lawmakers and national security officials who view advanced AI as a strategic technology with implications for economic competitiveness, military planning and cybersecurity.
Others, however, caution against equating benchmark performance with overall leadership.
Developing frontier AI systems requires far more than strong model outputs. Access to vast computing resources, advanced semiconductor chips and enormous datasets remains essential. Many analysts believe these areas continue to favour American firms, particularly as US export restrictions limit China’s access to the most advanced hardware.
The cybersecurity dimension
The stakes extend beyond commercial competition.
This week, intelligence and cybersecurity agencies from the Five Eyes alliance, which includes the United States, United Kingdom, Canada, Australia and New Zealand, warned that advanced AI models could dramatically reshape cyber operations in the near future.
Officials increasingly fear that highly capable AI systems could accelerate vulnerability discovery, automate cyberattacks and improve digital espionage capabilities. Those concerns become more urgent if rival nations gain access to comparable technologies.
Former Facebook security chief Alex Stamos recently warned against underestimating China’s capabilities, arguing that it would be a mistake to assume the best AI systems are necessarily being developed in the United States. He also suggested that public disputes within Washington over AI policy could ultimately benefit geopolitical competitors.
At the same time, China’s embrace of open-source AI may prove strategically important. By making advanced models more accessible, Chinese developers could attract businesses and governments looking for alternatives to costly proprietary systems produced by American companies.
The risk of being caught off guard
Perhaps the biggest concern among researchers is not what Chinese models can do today but what they may be capable of tomorrow.
Security experts increasingly use frontier AI models to study how rapidly abilities such as software vulnerability detection, persuasion and social engineering are improving. Some worry that restricting access to the most advanced Western models could make it harder for defenders to understand emerging threats if rival systems continue to advance.
Yet there is still no consensus that China’s rise will automatically translate into AI dominance. American companies continue to lead in several critical areas, while some cybersecurity firms claim their own specialised systems already outperform certain frontier models on specific tasks.
What is clear, however, is that the competitive landscape is changing. Each successive model release from China is making it harder to dismiss the country’s ambitions as merely aspirational. Whether the gap is measured in years, months or even weeks, the race for AI leadership is increasingly becoming a contest between two technological superpowers.
