Chinese AI models close gap with Anthropic and OpenAI as GLM-5.2 tops open-weight rankings

Chinese Z.ai's latest model tops AI ranking charts amid Anthropic Fable 5 ban -- blacklisted China firm's popular open-weight GLM-5.2 AI model powered by Huawei silicon

Z.ai's open-weight model leads the accessible field on domestic chips. On June 12th, the U.S. Commerce Department issued an export-control directive barring Anthropic from supplying Fable 5 or Mythos 5 to any foreign national, forcing the company to disable both models worldwide. The next day, Beijing-based Z.ai, formerly Zhipu AI, began rolling out GLM-5.2, an open-weight model it released under a permissive MIT license. The new model was purportedly trained entirely on Huawei Ascend chips with no Nvidia hardware. Within a week, GLM-5.2 had climbed to the top of the openly available leaderboards, Z.ai's market value had passed HK$1 trillion (about US$128 billion), and the most capable model many users outside the U.S. could legally access was a free download from a company that sits on Washington's trade blacklist. Trailing Anthropic GLM-5.2's results are both strong and uneven, taking first place on Design Arena's human-preference coding board, finishing roughly 10 Elo points ahead of Fable. It also ranks as the top openly available model on Artificial Analysis's Intelligence Index v4.1, where its score of 51 sits ahead of MiniMax-M3, DeepSeek V4 Pro, and Google's Gemini 3.1 Pro Preview. On the SWE-bench Pro, it scored 62.1, compared to GPT-5.5's 58.6. In terms of longer work, such as Code Arena's front-end board, the picture changes somewhat, with GLM-5.2 landing second behind Fable 5. On Artificial Analysis's AA-Briefcase test, which scores multi-week knowledge tasks built from thousands of fragmented inputs, Fable 5 led with 1,587 Elo, followed by Opus 4.8 at 1,356, and GLM-5.2 in third place at 1,266, before the export ban took Fable out of contention. It also trails on raw terminal work, scoring 81.0 on Terminal-Bench 2.1 against Opus 4.8's 85.0 and GPT-5.5's 84.0, while clearing Google's Gemini 3.1 Pro at 74.0. GLM-5.2 holds the top accessible position today, partly because the models that beat it on these benchmarks are largely an Anthropic pair, and Fable is now switched off. No Nvidia GLM-5.2's training stack is a slap in the face of Washington's efforts to curtail Chinese model development. Z.ai has been on the U.S. Entity List since January 2025, cutting it off from Nvidia's H100, H200, and B200 accelerators, and it says the GLM-5 family was trained on roughly 100,000 Huawei Ascend 910B processors using the MindSpore framework, with no Nvidia silicon at any stage. The export controls on advanced AI chips were designed to keep this kind of result out of reach, but they've evidently failed to do so. That said, the Ascend 910C sits at roughly 60% of an Nvidia H100's inference performance, per a December report from the Council on Foreign Relations, with a wide gap on efficiency and cluster scale. The same report projects that by as early as next year, the best U.S. chips could be more than 17 times more powerful than Huawei's top parts. At the same time, Huawei has claimed that a 1,000-chip Ascend cluster handled full-parameter post-training of DeepSeek's V4. If true, this shows that Chinese domestic silicon can now carry training-class jobs, just not at Nvidia's per-chip throughput or scale. So, while GLM-5.2 demonstrates that a frontier-class open model can be produced on a fully domestic stack, it doesn't demonstrate that the chips underneath have caught up with Nvidia; model parity =/= hardware parity. The Fable 5 shutdown Anthropic released Fable 5 to the public on June 10th, a safety-restricted build of its Mythos 5 model designed to block the cyber and bio capabilities of the underlying system. Just two days later, the Commerce Department suddenly and unexpectedly ordered access to be pulled for all foreign nationals, including Anthropic's own non-citizen staff, after officials cited a technique for bypassing Fable 5's safeguards. Anthropic said in an announcement following the restriction that the jailbreak it understood to be at issue was narrow rather than universal, surfaced only previously known minor vulnerabilities, and produced behavior also obtainable from other public models, including OpenAI's GPT-5.5. The company said in its statement that it believes the order rests on a "misunderstanding" and is working to restore access. But because the directive covered all foreign nationals, Anthropic had no way to keep the models live for U.S. users alone and disabled them for everyone. Meanwhile, GLM-5.2's MIT license lets anyone download, fine-tune, and self-host its weights, which is the basis for calling it a freely available model. Running it, however, is a separate matter: the model carries around 744 billion total parameters, 40 billion of them active per token, with a one-million-token context window. That's no small footprint and calls for enterprise GPU clusters or high-memory workstations -- it's not something you're ever going to get running on a desktop -- and throughput drops sharply once context runs past tens of thousands of tokens. The most practical way to use GLM-5.2 is via the API, where Z.ai prices the model at about $1.40 per million input tokens and $4.40 per million output, against $5 and $25 for Claude Opus 4.8, or $10 and $50 for Fable 5. On the AA-Briefcase runs, Fable 5 averaged $31 per task to GLM-5.2's $2.40, a roughly 13-times spread that holds even where Fable scored higher. The market moved fast with GLM-5.2's release. Z.ai, which is listed on the Hong Kong exchange as Knowledge Atlas Technology, saw its shares jump as much as 42% intraday on June 22nd to HK$2,980, carrying its market capitalization past HK$1 trillion. Founder Tang Jie has said publicly that a Chinese model matching Fable 5 will arrive sooner than the first-quarter timeline Elon Musk recently floated. There's a nearer date, too. On July 8th, the lock-up on Z.ai's first cornerstone investors expires, freeing a large block of shares to trade, which will give the GLM-5.2 rally its first real test.

A new, inexpensive Chinese AI model is catching up with Anthropic, OpenAI on their home turf

BEIJING/BENGALURU, July 2 (Reuters) - Since DeepSeek shocked markets early last year with its cheap but powerful AI model, global consumers have been faced with a choice: Chinese offerings with lower prices and less capability or OpenAI or Anthropic, which have poured billions into development. A model called GLM-5.2, launched last month by Beijing-based startup Z.ai, may finally be closing that gap in terms of Western interest. GLM-5.2 has Silicon Valley buzzing with its coding and agent capabilities, or the ability to execute complex tasks with minimal prompting, that almost rival leading U.S. offerings at a fraction of the cost, in what some experts are calling a "mini DeepSeek moment." It has quickly climbed the usage charts on third-party AI developer platforms like OpenRouter, where it now ranks above Anthropic's models, while executives from cloud data platform Snowflake's CEO Sridhar Ramaswamy to venture capitalist Marc Andreessen have lauded its abilities. "We now have a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic," said David Sacks, U.S. President Donald Trump's former AI czar, last week before Washington lifted curbs on Anthropic's Fable and Mythos models on Tuesday. Those capabilities have put Z.ai's GLM-5.2 model at the heart of a growing debate about whether China is finally catching up to the U.S. in the AI race, as technology executives warn that Washington's unpredictable ⁠regulation of the industry risks hampering its lead in the frontier technology. "It is just a tick below Opus 4.8 (from Anthropic) and right up there with GPT 5.5 (from OpenAI)," Sacks said of GLM-5.2 on the All-In podcast, adding that "we cannot afford to do things that slow our companies down." The Anthropic curbs and the delayed public rollout of OpenAI's latest GPT-5.6 model have fueled global demand for the Chinese model, some experts said. "The international developer community is increasingly aware that relying solely on proprietary, U.S.-based API models carries significant risk," said Brian Tse, founder and CEO of Concordia AI, a Beijing-based consultancy focused on AI safety. GLM-5.2's positive global reception also suggests increased interest in cheaper open-source development because businesses are getting stung by the rising and often unpredictable costs of using AI to complete tasks, as closed-source agentic AI tools consume more tokens, the units used to measure AI usage. Z.ai, also known as Zhipu AI, declined to comment. Anthropic and OpenAI did not immediately respond to requests for comment. GLM-5.2 currently holds fifth place on Artificial Analysis' large language model (LLM) intelligence leaderboard, which ranks performance across a range of benchmarks designed to measure overall capability, including reasoning and coding skills. And it is in the second spot on Code Arena's front-end coding rankings, measuring how well models generate websites and front-end applications, while operating at roughly a sixth of the cost of closed U.S. frontier models like Claude and the GPT series. Z.ai has not disclosed how ⁠much it spent to develop GLM-5.2. In a reply to Elon Musk on X last month, Z.ai founder Tang Jie said that the Chinese startup could produce a model on par with Anthropic's Fable before the first quarter of next year. "The shift GLM-5.2 brings is that the open-source model has become a plug-and-play, out-of-the-box product," said Tiezhen Wang, former APAC lead at Hugging Face, a startup that serves as a hub for developers tinkering with open-source models. "You just deploy the model and without doing any complex fine-tuning systems, it is in a highly usable, ready-to-use state. This drastically lowers the barrier to entry for open-source adoption." CONVINCING AMERICAN BUSINESSES One major hurdle to GLM-5.2's large-scale adoption remains data security concerns that have limited use of Chinese models by U.S. enterprises, particularly in regulated industries like ⁠banking and cybersecurity. The migration and upgrading of enterprise AI systems typically takes several months, Wang said. "I have seen some discussion among European companies about whether it could be used in enterprise settings," said Wei Sun, principal AI analyst at Counterpoint Research. "In the EU and U.S., some clients, partners and regulated industries may simply be unwilling to accept Chinese models in their AI stack, regardless of technical performance or price." A report earlier this year by non-profit RAND, opens new tab, based on website traffic data across 135 countries, found that Chinese LLMs' global ⁠market share jumped to 13% from 3% in the two months after DeepSeek launched its R1 model in January last year. The release sparked a global tech selloff because it contrasted DeepSeek's low cost with massive AI infrastructure spending elsewhere. China's LLM usage gains were most pronounced in developing countries and those with close political and economic ties to Beijing. Some experts said concerns about the safety of Chinese AI models were overblown, arguing that running them on U.S. cloud providers or ⁠on a company's own servers ensured data security. While major corporations are slow to migrate, tech startups and small- and medium-sized enterprises are moving much faster. "Developers tend to care less about where a model comes from than whether it works, how much it costs and whether they can deploy or access it reliably," said Poe Zhao, China tech analyst and founder of the Hello China Tech newsletter. "The likely pattern is partial routing, not overnight replacement of OpenAI or Anthropic. So yes, it is a mini DeepSeek moment but in a narrower, developer-centric sense." Reporting by Laurie Chen in Beijing and Aditya Soni in Bengaluru; Editing by Eduardo Baptista and Thomas Derpinghaus Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Artificial Intelligence Laurie Chen Thomson Reuters Laurie Chen is a China Correspondent at Reuters in Beijing, whose coverage focuses on the nexus of frontier technology, strategic emerging industries and geopolitics. She has reported on China for almost a decade, having previously covered China's government, defence, security and foreign policy. She has broken multiple global scoops on U.S.-China relations and the trade war 2.0, elite Chinese politics and diplomacy. She is particularly interested in Chinese frontier AI, tech and industrial policy, semiconductor supply chains, robotics, aerospace and grand strategy.

White House AI crackdown opens door for Chinese model makers to close gap

The Chinese conundrum lands as much of corporate America is moving from an era of tokenmaxxing to a focus on efficiency. The Trump administration's crackdown on Anthropic's leading artificial intelligence models is looking like a gift to the country's top adversary in the AI race: China After a two-week shutdown due to an export control directive, Anthropic was allowed by the White House on Friday to release its powerful Mythos 5 model to some companies and federal agencies, though its Fable 5 model remains off the market. OpenAI, meanwhile, also said on Friday it would limit the rollout of its GPT 5.6 models following a government request. The two leading U.S. AI model developers have been in a race against each other and tech giants like Google to develop the most advanced technology, with the U.S. government opening the door to speedy AI development by limiting regulatory hurdles. Doing otherwise, according to many tech execs and Trump administration officials, would restrict domestic AI to the benefit of China, which has been rapidly catching up to the U.S. But as Anthropic adheres to the U.S. government's national security concerns, Chinese companies are launching models that rival frontier labs in some capabilities. According to researchers, Zhipu's GLM 5.2, released earlier this month, can perform on par with top U.S. labs on some cyber benchmarks, even equaling Mythos' capabilities. "Many smart people/AI insiders are saying GLM-5.2 is the first Chinese AI model to match and often beat the American big lab public AI models with no compromises," wrote venture capitalist Marc Andreessen, in a post on X over the weekend. "Incredible timing given current events." Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology, called the recent developments "a pretty good wake-up call," and Jefferies strategist Christopher Wood wrote in a report to clients, citing industry sources, that GLM 5.2 "is almost equal to Anthropic as a competitor for the corporate market and is just one quarter of the cost in terms of cost per token." Even former Trump crypto and AI czar David Sacks, who's been a vocal critic of Anthropic's approach to AI safety, wrote a cryptic post on X above a screenshot of a Wall Street Journal headline stating that China has matched Anthropic in cybersecurity. "A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy, and pro-export," Sacks wrote. "President Trump was exactly right; we deviate from that strategy at our peril." Representatives from Anthropic, OpenAI and the White House didn't respond to requests for comment. The Chinese conundrum lands just as much of corporate America is moving from an era of so-called tokenmaxxing, or allowing developers to spend on AI without restraint, to a focus on efficiency and return on investment. That also plays into China's hands. Earlier this month, Flo Crivello, CEO of AI startup Lindy, switched his company off Anthropic's Claude models and moved 100% of its traffic to DeepSeek, a Chinese company that makes cheaper, open-weight alternatives. "We did it, and you could see that cost curve go down, like, crash to the ground," Crivello told CNBC for a story published last week. Chinese AI developers are reaching U.S. users with ease because of how simple it is for a company to download open-weight models and run them on their own servers without relying on a third-party cloud. "With the open-weight models, it's kind of the Wild West," said Travis Lanham, co-founder of AI security startup Armadin, which is experimenting with GLM 5.2 and another Chinese model, Kimi K2.7 from Moonshoot AI. Lanham said the models are showing improved capabilities for cybersecurity use cases like analyzing reconnaissance data and creating customer exploit code. Whether U.S. authorities will allow that to continue is an emerging question in policy circles due to how the two largest economies handle each other's sensitive hardware. For years, the U.S. government has gone to great lengths to keep cutting-edge AI innovation out of China's hands through U.S. export controls on AI chips from Nvidia and Advanced Micro Devices. The U.S. also banned American companies from using Huawei equipment because of national security fears. Last year, the U.S. cleared Nvidia's H200 chip -- the same model used by U.S. companies -- for export to the China region. But Nvidia said earlier this year that it had yet to generate any revenue from the chips and that it didn't know whether China would allow imports of its products. As for GLM 5.2, Tesla and SpaceX founder Elon Musk wrote in a post on X that the model would probably reach Fable levels by the first quarter, responding to a question from a user asking when that milestone would be reached. Zhipu founder Jie Tang, in a reply to Musk, wrote, "won't take that long." It's not just niche brands turning to Chinese models. Major companies like Shopify and Airbnb have touted the benefits of Alibaba's Qwen 3 for scaling AI features. Coinbase CEO Brian Armstrong wrote on X last week that his company is utilizing open-weight models such as GLM 5.2 and Kimi 2.7, allowing it to cut nearly half its AI spending despite greater token use. Cybersecurity is the primary concern for many industry experts. Some open-weight models can already automate many stages of a cyberattack, and Hed Kovetz, CEO of industry startup Silverfort, worries they're only months away from running an entire operation. "If the U.S. government does not let the industry take advantage of this opportunity to get ready, then when the Chinese models reach a similar level, no one will be prepared," he said. -- CNBC's Deirdre Bosa and Ashley Capoot contributed to this report. Choose CNBC as your preferred source on Google and never miss a moment from the most trusted name in business news.

Chinese A.I. Models Gain Ground on Anthropic and OpenAI

Cade Metz reported from San Francisco, Karen Weise from Seattle and Meaghan Tobin from Taipei, Taiwan. Two weeks ago, the artificial intelligence company Anthropic shut down its two most powerful A.I. systems after an unexpected demand from the U.S. government to cut access to it. Days later, a Chinese start-up, Z.ai, released an A.I. model that is nearly as powerful as Anthropic's models, Fable and Mythos. But Z.ai's new technology costs much less to use, and no one in the United States was putting restrictions on it. It quickly landed on a closely watched leaderboard of the world's 10 most popular models. Z.ai is on the cutting edge of a wave of powerful but inexpensive A.I. from China that is challenging the lock that OpenAI, Anthropic and Google have had on the industry. Six of the models now on the A.I. leaderboard were developed in China. Z.ai's new model, GLM-5.2, arrived just as U.S. businesses realized that they had to find ways to cut down on how much they were spending on A.I. It also landed when executives in Silicon Valley were becoming worried that the Trump administration was leaning toward regulating the technology. "With Fable restricted, the gap between the U.S. and China is very slim," said Rehaan Ahmad, a co-founder of the Silicon Valley start-up alphaXiv, who has been using Z.ai's new model for more than a week. The Chinese models still face two big hurdles to widespread use in the United States: concerns about their ties to the Chinese government and complaints that Chinese companies have unfairly used American technology to build these cheaper models. But their low cost is winning converts. About 18 months ago, the Chinese start-up DeepSeek shocked Silicon Valley when it demonstrated that it could build effective A.I. far more affordably than many of its American counterparts. Z.ai is doing something similar. When performing certain tasks, GLM-5.2 costs about an eighth as much as Anthropic's Claude Opus 4.8, which came out shortly before Fable and Mythos, according to OpenRouter, a start-up that runs the A.I. leaderboard. Like most top-performing Chinese models, GLM-5.2 is open source software, which means anyone can use and modify it for free. That makes it much cheaper to use, even if it is not quite as powerful as what American companies have created. "Do you need to drive a Ferrari everywhere?" asked Vivek Ramaswami, a start-up investor at Madrona Venture Group. "Probably not." Z.ai did not respond to requests for comment. GLM-5.2 is particularly good at generating computer code and powering A.I. agents, digital assistants that can use other software to perform tasks. Z.ai's technology is now the third most widely used in the world for A.I. tasks, said Anastasios Angelopoulos, chief executive of ArenaAI, which tracks millions of A.I. users. The largest cloud computing providers, including Microsoft and Amazon, already offer access to some systems from Z.ai, DeepSeek, MiniMax and other Chinese start-ups. Microsoft has also considered adding the latest DeepSeek model as an option to power one of its own products, which now runs on technology from Anthropic and OpenAI, two people familiar with the deliberations said, speaking on the condition of anonymity because they were not authorized to discuss them publicly. The talks were reported earlier by Axios. Microsoft, Anthropic and OpenAI declined to comment. Some software developers are reluctant to use the A.I. system that Z.ai offers from computers in China, because they worry about sharing data with the company or with the Chinese government. They are also wary of China's efforts to censor its A.I. systems or running afoul of U.S. export restrictions. Z.ai was added to the Commerce Department's trade blacklist in 2025. Corporate filings show that several of the company's shareholders are controlled by a Chinese government agency that supervises the country's defense industry. Companies can still use the model without sending data back to China in violation of U.S. export rules, as long as they are careful about how they set up their systems, said Wei Chen, chief legal officer at Infoblox, a network security company. "The Chinese models do not have the same restrictions if you host them yourself or you go through another provider," Mr. Ahmad of alphaXiv said. "Right now, there are more restrictions on models from Anthropic." After DeepSeek's release in 2025, governments around the world passed regulations limiting its use because of data security concerns. But so far, GLM-5.2 has not raised similar alarms, Ms. Chen said. Anthropic and OpenAI have accused Chinese companies of improperly harvesting data from their A.I. systems to accelerate the development of the Chinese technology. On Wednesday, Anthropic sent a letter, viewed by The New York Times, to Senators Tim Scott, Republican of South Carolina, and Elizabeth Warren, Democrat of Massachusetts, accusing the Chinese tech giant Alibaba of "brazenly" and "illicitly" trying to copy its technology through 24,000 fraudulent accounts. Alibaba declined to comment. When Mustafa Suleyman, the head of Microsoft's A.I. lab, unveiled a suite of new models this month, he emphasized that they had been built from scratch on data that the company had commercially licensed. "That means that you can put it into production in a very trustworthy way with complete confidence," he said. (The Times has sued OpenAI and its partner, Microsoft, accusing them of copyright infringement of news content related to A.I. systems. They have denied those claims.) Using data from one system to train another -- a process called distillation -- is common in A.I. development. But the Anthropic and OpenAI terms of service forbid anyone to surreptitiously harvest data for distillation. It is not clear whether Z.ai used distillation in the development of its technology. But distillation alone cannot build a top A.I. system. That requires several other complex techniques as well, said Charles O'Neill, head of model training at Baseten, a company that sells access to GLM-5.2. "This narrative that all of the capabilities of these models are coming from Anthropic is not as true as people say it is," Mr. O'Neill said. Chinese A.I. start-ups can offer their models as an open-source technology at far lower prices in part because the industry has benefited from years of support from the Chinese government, which views A.I. as a critical engine of economic growth. Many executives have said U.S. companies should not open-source their technology because it could be used in harmful situations. But other experts argue that if regulators stifle open-source technology in the United States, China will gain a significant edge. Because China produces most of the top-performing open-source systems, they say, U.S. developers will build their software atop those technologies. In the long run, that could put China at the heart of A.I. development. Some argue that Chinese systems will always trail the top U.S. models because U.S. export controls limit the flow of the specialized computer chips needed to train A.I. technologies. Z.ai and other Chinese start-ups spend millions for access to chips in data centers outside China. Z.ai's filings in Hong Kong show that in the first half of 2025, the company spent more than seven times its revenue on expenses that essentially boiled down to fees for such computing services. Still, experts estimate that China is just six months or less behind the American companies. "There had been this speculation that the export controls would eventually bite and the gap would widen between American frontier models and their Chinese models, but GLM is pushing things in the other direction," said Jeffrey Ding, an assistant professor at George Washington University who specializes in emerging technology and international relations. And with Fable and Mythos sidelined, many businesses have realized the importance of having alternatives. "There is a bit of apprehension at large organizations about loyalty," said Justin Summerville, who runs data analytics at OpenRouter. "Who knows what the top model will be in three weeks?"

A cheap Chinese AI model is closing in on Anthropic and OpenAI

Z.ai's GLM-5.2 has climbed to fourth on a leading intelligence benchmark, running on domestic Chinese chips at a fraction of Western frontier pricing Chinese startup's latest AI model has landed fourth on one of the industry's most closely watched intelligence rankings, and it costs a fraction of what Anthropic or OpenAI charge for comparable performance. GLM-5.2, released last month by Beijing-based Z.ai, has become the talk of Silicon Valley for coding and agentic capabilities that edge close to the leading American systems, prompting comparisons to DeepSeek's market-rattling debut in 2025. According to Artificial Analysis, GLM-5.2 scores 51 on the firm's Intelligence Index v4.1, placing it fourth overall and first among all open-weight models, ahead of MiniMax-M3, DeepSeek V4 Pro and Kimi K2.6. On Code Arena's front-end coding leaderboard, the model's Max tier holds second place, ahead of Anthropic's Claude Opus variants. The pricing is the part that has unsettled rivals. Z.ai charges $1.40 per million input tokens and $4.40 per million output tokens on its first-party API, and third-party hosts list it even lower. Multiple outlets have pegged that at somewhere between a fifth and a seventh of what Claude Opus or GPT-5.5 cost per output token, though the exact ratio shifts depending on which provider and tier is being compared. Either way, it undercuts the closed American frontier by a wide margin. David Sacks, who served as the White House's AI czar under Donald Trump before returning to the private sector, said the model was "as good as the currently available models from OpenAI and Anthropic," describing it as sitting just below Anthropic's Opus 4.8 and roughly level with OpenAI's GPT-5.5. That is a striking claim from someone who, only weeks earlier, was estimating the US lead over Chinese labs at six to nine months rather than conceding rough parity outright. Beyond the headline intelligence score, GLM-5.2 also tops Artificial Analysis's GDPval-AA v2 metric, which measures real-world agentic work rather than abstract reasoning puzzles. There it scores 1,524, ahead of MiniMax-M3 and DeepSeek V4 Pro, and close enough to GPT-5.5 on its highest reasoning setting that the two are effectively tied. Analysts have flagged one trade-off worth noting: the model uses noticeably more output tokens per task than its open-weight peers, which eats into some of its cost advantage in practice even as the sticker price stays comparatively low. GLM-5.2 also arrives with a detail that matters more than the benchmark scores: it was trained and runs on domestic Chinese silicon, reportedly a cluster of roughly 100,000 Huawei Ascend 910B processors, without Nvidia, AMD or Intel hardware at any stage. That is a direct answer to Washington's chip restrictions, which have aimed to slow Chinese AI progress by cutting off access to the most advanced processors. The model's weights are released under an unrestricted MIT licence, meaning any company can download, modify and run it locally for the cost of electricity alone. The timing has helped Z.ai's case. Washington's order forcing Anthropic to suspend Fable 5 and Mythos 5 for foreign users, combined with the Trump administration's request that OpenAI stagger the rollout of its own next model, has left a gap in the market that a free, capable, downloadable alternative is well placed to fill. That backdrop sits inside the broader story of tightening US export controls on advanced chips, which Chinese labs have spent the past two years working around rather than waiting out. Z.ai's listed entity has felt the effect directly. Shares in the Hong Kong-listed company have risen sharply since its stock market debut in January, with trading spikes around GLM-5.2's release pushing the stock up several multiples over its IPO price. A successor model, GLM-5.5, is expected in August, though Z.ai has not confirmed a firm date.

After Anthropic shutdown, China's Z.ai closes frontier gap as it plans dual listing

BEIJING, June 25 (Reuters) - Chinese AI startup Z.ai (2513.HK), opens new tab said on Thursday it plans to use domestic listing proceeds to fund a quest for artificial general intelligence, after its latest model scored close to leading U.S. models from Anthropic and OpenAI on public benchmarks. Z.ai's flagship GLM-5.2 model, released a day after Anthropic disabled worldwide access to its most advanced models, stunned global users after its performance benchmarks narrowly trailed leading closed-source models. "Our mission is to obtain AGI, so right now our focus is how to improve our model to achieve the upper bound of intelligence. So all these resources are helping us," said Qinkai Zheng, technical lead of the firm's CodeGeeX team. The accompanying surge in investor interest sent shares rallying more than 2,000% from their blockbuster Hong Kong debut in January, to cross a threshold of HK$1 trillion ($128 billion) in market capitalisation this week. "This model is comparable to the top closed models," Zheng told reporters at its head ⁠office in Beijing. "It's the first time that an open-source model really delivers very solid coding and agent performance that can compare with the leading proprietary AI companies like Anthropic and OpenAI." GLM-5.2 now holds fourth place on Artificial Analysis' LLM intelligence leaderboard and the second spot on Code Arena's front-end coding leaderboard, operating at roughly a sixth of the cost of closed U.S. frontier models. The sudden "unplugging" of closed U.S. frontier models has unleashed severe anxiety among global allies such as Canada and France, whose leaders heavily criticised over-reliance on U.S.-controlled AI infrastructure at a G7 summit last week. For the first time a Chinese open-source AI model has come close to bridging the frontier gap with heavily-funded Western AI labs, after previously surpassing U.S. open-source models such as Google's Gemma and Meta's Llama series despite constraints on computing power. Analysts previously estimated that the performance capabilities of top Chinese AI models were four to six months behind leading U.S. models. Z.ai, also known as Zhipu AI, said this month it plans a dual listing in the ⁠commercial hub Shanghai but did not disclose how much it aimed to raise. SPECIALISING IN CODING, COMPLEX TASKS Specialising in coding and complex long-horizon tasks, the model has 750 billion total parameters and a massive 1-million token context window. It was released with inference adaptation to a wide variety of domestic chip infrastructure users, including Huawei Ascend clusters, the company said in a blog post. Since February, Z.ai's GLM-5 series has been adapted to run on domestic semiconductors after the U.S. tightened China's access to advanced Nvidia chips. "We are trying our best to improve our infrastructure and to ⁠make the model more efficient on different kinds of chips," said Zheng, without disclosing whether it was trained on domestic or foreign chips. Despite fierce domestic price competition, Z.ai has successfully hiked prices for its frontier models multiple times this year, reflecting its strong enterprise adoption in China, where it is also widely used in public sector procurement. In a post on X, co-founder Tang ⁠Jie emphasised Z.ai's commitment to open-source and described the abrupt pulling of Anthropic's Fable and Mythos models as "deeply regrettable". JP Morgan projected the firm's revenue to balloon by this year and for it to turn profitable in 2028 - but Z.ai only earns a fraction of the revenue of its U.S. counterparts, stock exchange filings show. "We are trying to ⁠lower the cost, but because the demand is too large, maybe in the future we will still need to increase the price," said Zheng. "But we want the model accessible to everyone." Zheng said that the company would target long-horizon tasks and self-evolving autonomous agent systems in future models. Its next GLM-5.5 model is expected to be released in August and could become the next key milestone in Chinese frontier AI. ($1=7.8397 Hong Kong dollars) Reporting by Laurie Chen; Editing by Clarence Fernandez Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Disrupted Laurie Chen Thomson Reuters Laurie Chen is a China Correspondent at Reuters in Beijing, whose coverage focuses on the nexus of frontier technology, strategic emerging industries and geopolitics. She has reported on China for almost a decade, having previously covered China's government, defence, security and foreign policy. She has broken multiple global scoops on U.S.-China relations and the trade war 2.0, elite Chinese politics and diplomacy. She is particularly interested in Chinese frontier AI, tech and industrial policy, semiconductor supply chains, robotics, aerospace and grand strategy.

China's Zhipu is closing in on top U.S. AI models with Anthropic and OpenAI held back

Zhipu's GLM 5.2 artificial intelligence model landed last week with the kind of Silicon Valley buzz that followed DeepSeek's launch last year. The Chinese AI startup's new open source model has blown past every other open release and now sits within a percentage point of Anthropic's Opus 4.8 on one closely watched agentic benchmark, at roughly a fifth of the cost. Developers are piling in, with OpenRouter token traffic climbing faster than it did after DeepSeek's V4 launch in April. And unlike DeepSeek, which the market eventually dismissed as a one-off chatbot shock, GLM 5.2 is strong at agentic work like planning, coding, testing, and looping, the kind of work that enterprises are racing to automate. Companies hit by unexpectedly high AI spend on tokens, the measure of the data processed and generated by AI models, are now asking how to get the most for their money. The most important metric is becoming intelligence per dollar, making Zhipu's cheap-but-good-enough model an attractive answer. "I've been consistently surprised by how quickly the open source has caught up," Gabe Pereyra, co-founder of Harvey, told CNBC. "GLM 5.2, you're seeing the first model where it's really competitive with some of these closed-source frontier models." But the real story underneath the price tag is the open source AI moment. GLM 5.2 is free to download, fine-tune, and run on an enterprise's own servers, putting pricing pressure on frontier labs at the same time that access looks shaky. Anthropic had to pull its Fable Mythos-class model after an order by the Trump administration, and OpenAI announced Friday that it is limiting its GPT 5.6 models because of a government request. The federal oversight has made a model that no one can revoke increasingly look like the safer bet.

How Chinese AI models are closing the gap on Anthropic and OpenAI

The temporary sidelining of Anthropic may be over, but a Beijing-based alternative made hay while the sun shone. In mid-June, artificial intelligence company Anthropic shut down its two most powerful AI systems after an unexpected demand from the US government to cut access to them. Days later, a Chinese start-up, Z.ai, released an AI model that is nearly as powerful as Anthropic's models, Fable and Mythos. But Z.ai's new technology costs much less to use, and no one in the United States was putting restrictions on it.

A new, free Chinese AI model is rattling U.S. tech investors

China's Zhipu AI says its newest model can find software security bugs as well as Anthropic's most tightly restricted system. The claim landed this week with no benchmark paper behind it, just a viral post. That hasn't stopped it from rattling a market already nervous about what's propping up American AI valuations. The system Zhipu is comparing itself to, Claude Mythos, isn't available for outsiders to test. Washington ordered Anthropic to cut off foreign access to Mythos and its public sibling, Fable 5, on June 12, citing national security. That means Zhipu's claim can't be independently verified, and may never be in its current form. A real benchmark sits underneath the unverified one Strip away the cybersecurity headline and a harder number remains. Zhipu's open model, GLM 5.2, landed last week with the kind of buzz that followed DeepSeek's debut a year earlier. It has since pulled ahead of every other open release, sitting within a single percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost, a CNBC report found. Developers have noticed: OpenRouter token traffic for GLM 5.2 is climbing faster than it did after DeepSeek's V4 launch in April, per the same report. That comparison matters because DeepSeek's moment faded once the market decided it was a one-off chatbot scare. GLM 5.2 is built for agentic work instead, planning, coding, testing, and looping through tasks on its own, which is exactly the kind of labor enterprises are trying to automate right now. As AI token spending has climbed past what many companies budgeted for, intelligence per dollar has become the metric that matters, and a cheap, good-enough open model is a direct answer to that pressure. "I've been consistently surprised by how quickly the open source has caught up," Gabe Pereyra, co-founder of Harvey, told CNBC. hapabapa / Getty Images Open and closed models are now solving different problems The distinction underneath all of this is simple but easy to miss. Closed models like Opus 4.8 and Mythos live entirely on Anthropic's servers, accessed through an API the company, or a regulator, can switch off at will. Open models like GLM 5.2 ship their weights under a license, MIT in this case, that lets anyone download, modify, and run the system on their own hardware. That difference used to be mostly philosophical. It's now a procurement decision. A company self-hosting GLM 5.2 isn't exposed to a future export order, a pricing change, or an outage at someone else's data center, though it does take on the engineering cost of running the model itself. Hosting through Zhipu's own cloud API instead reintroduces a version of that dependency, since usage would fall under Chinese law rather than a self-hosted server's jurisdiction. Access, not just capability, is now part of the valuation math Here's what should bother U.S. AI investors more than any single benchmark: The National Security Agency lost its own access to Mythos as a result of the export order, Nextgov reported, even though NSA analysts had been using it to find vulnerabilities in classified systems. A model good enough for the government to rely on got pulled before the government finished relying on it. OpenAI separately moved to limit GPT-5.6 to "trusted partners" at the U.S. government's request the same week. A model nobody can revoke is starting to look like the safer bet to enterprises planning years out, regardless of which one currently scores higher on a leaderboard. Big news: Mythos is partially back The test arrived sooner than expected. The U.S. government cleared Anthropic to restore Mythos 5 access on Friday, but only to roughly 100 vetted organizations focused on critical infrastructure and cyber defense, NBC News reported. Fable 5, the public-facing version most enterprises actually used, stays offline. That distinction is the real signal. Commerce Secretary Howard Lutnick could amend the approved list at any time. For the roughly 100 firms back in, the tool is faster than building the same defenses from scratch. For everyone else, the lesson Zhipu has been selling all month just got reinforced by the U.S. government itself: A model you can't self-host is a model someone else can take away. That's the deeper problem for Anthropic and OpenAI's combined $1.8 trillion-plus valuation. A free, downloadable alternative closing the gap on cost and agentic performance was already a hard sell. A premium model whose access list a Commerce Secretary can edit by letter, even temporarily, makes the case harder still. The Arena Media Brands, LLC THESTREET is a registered trademark of TheStreet, Inc. This story was originally published June 30, 2026 at 2:03 AM.

China's GLM-5.2 AI Model Gains Ground as US Rivals Face Access Limits

Developers can download the model weights, adjust the system and run it on private infrastructure. This gives companies more control over data, security and operating costs. By contrast, many leading US models depend on paid cloud access and provider-controlled application programming interfaces. Early tests have placed GLM-5.2 near leading proprietary models on selected coding and agent benchmarks. Reports have also shown lower token prices than several frontier systems. These comparisons have increased interest among companies that want capable models without the highest usage fees. Still, benchmark results require caution. Scores can change with prompts, tools, computing settings and test rules. Zhipu published performance data for the model, but users will need broader independent testing across real business tasks. Harvey co-founder Gabe Pereyra said, "I've been consistently surprised by how quickly open source has caught up." His comment reflects early interest, not proof that GLM-5.2 leads every category. Meanwhile, OpenRouter lists GLM-5.2 with a one-million-token context window and prices below several premium closed models. Traffic data has also shown rapid developer adoption. Yet high usage does not always show strong reliability, security or long-term support.