{"slug": "china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests", "title": "China's Z.ai Says Its New Model Nears Anthropic's Mythos 5 on Cyber Tests", "summary": "Z.ai, formerly Zhipu AI, has not publicly released GLM-5.3, and claims of its cyber benchmark wins over Anthropic's Mythos 5 are unverified; the company's verified flagship is GLM-5.2, an open-weight model with a 1 million token context window and MIT license. Meanwhile, Anthropic has kept its Mythos models gated under Project Glasswing, with selected partners finding over 10,000 high-severity vulnerabilities, and the US Commerce Department briefly restricted access to Fable 5 and Mythos 5 in June.", "body_md": "*Z.ai's real story isn't an unverified GLM-5.3 benchmark win over Anthropic. It's that Chinese open-weight models are moving fast while the most powerful cyber models in the US are being kept behind gates.*\n\nZ.ai, the company still widely known by its earlier Zhipu AI name, has not made GLM-5.3 into the public cyber victory lap described in the numbers floating around it. The checked record points somewhere less tidy and more useful for you: GLM-5.2 is the current hard anchor, Anthropic's Mythos line is still treated as sensitive cyber infrastructure, and the race between Chinese and US labs is now as much about access as raw scores.\n\nThat matters. If you're buying or building with these models, a leaderboard claim is not enough. You need the model card, the release note, the access terms, and the benchmark method. Without those, you're not looking at evidence. You're looking at noise with decimals.\n\n## The verified Z.ai model is GLM-5.2\n\nZ.ai's own June release materials describe GLM-5.2 as its flagship model for long-horizon work, with a 1 million token context window, stronger coding controls, and MIT-licensed weights. The Hugging Face card for zai-org/GLM-5.2 says the same thing, including the 1 million context claim and the open-weight release. That's the model you can actually check.\n\nThe CyberGym claim attached to GLM-5.3 is different. Searches for the alleged Reuters comparison, the 84.5 CyberGym score, the 54.4 ExploitBench score, and the six-hour ExploitGym task count did not turn up a Reuters article, a Z.ai release note, or a public benchmark page supporting those figures. So don't build a story around them.\n\nThere is still plenty here without inventing a finish line. GLM-5.2 is a large, open-weight Chinese model aimed squarely at coding agents and long-context engineering work. Its published materials say Z.ai changed the architecture with IndexShare, reusing an indexer across sparse attention layers to cut per-token work at long context. That is not a small product note. For developers, context length is only useful if the model can keep working across it without turning every request into an expensive crawl.\n\nPrice and access are the other half. Reuters reported in June that Knowledge Atlas Technology JSC, also known as Zhipu AI, planned to apply for a Shanghai STAR Market listing and change its English corporate name to Z.AI Co. Reuters also reported that the proposed A-share issue would represent between 2% and 8% of total share capital. A company pushing for public-market capital while shipping open models is not just doing research theater. It is trying to turn model access into a durable business.\n\n## Anthropic is playing a different game\n\nAnthropic's approach to Mythos is almost the mirror image. On April 7, the company announced Project Glasswing, giving selected partners access to Claude Mythos Preview for defensive cybersecurity work. Anthropic said the launch group included Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks, along with more than 40 additional organizations.\n\nIt kept the model gated. That is the point.\n\nAnthropic said Project Glasswing came with up to $100 million in model usage credits and $4 million in donations to open-source security organizations. In May, the company said roughly 50 partners had used Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities across important software. By June, Anthropic said it was expanding the program to about 150 more organizations across more than 15 countries, with security requirements before access.\n\nYou can see why governments are nervous. A model that helps defenders find old flaws in critical code can also help attackers move faster if it is released carelessly. The US Commerce Department briefly restricted access to Anthropic's Fable 5 and Mythos 5 models in June, according to Reuters and Anthropic's own statements, before the controls were lifted on June 30 after safeguards were added. Access to Mythos remained tied to vetted organizations through Glasswing.\n\nThat is not ordinary product gating. It is a sign that cybersecurity models have crossed into policy territory.\n\n## The benchmark race needs less theater\n\nThe weak version of this story is simple: Chinese model beats US model on one test. The stronger version is harder and more honest. Z.ai is proving that open-weight Chinese labs can move fast enough to keep pressure on closed US leaders. Anthropic is proving that the best cyber capability may not be something a lab can safely throw into a public API and call finished.\n\nBoth things can be true.\n\nFor founders, security teams, and developers, the practical call is blunt. Treat unsupported benchmark claims as marketing until the source is public and the method is clear. GLM-5.2 is worth watching because its release is documented, its weights are available, and its long-context pitch is specific. Mythos is worth watching because Anthropic is restricting access to it even while saying it found thousands of serious vulnerabilities.\n\nThe real competition is not one decimal point on CyberGym. It is who can make these systems useful without making them reckless.\n\n**Also read:** [Guangdong Taps Alibaba to Power Its AI and Semiconductor Push](https://startupfortune.com/guangdong-taps-alibaba-to-power-its-ai-and-semiconductor-push/), [Z.ai launches GLM-5.3, a coding model billed as ready for cyber defense](https://startupfortune.com/zai-launches-glm-53-a-coding-model-billed-as-ready-for-cyber-defense/), [OpenAI's Revenue Run Rate Tops $40 Billion Just Months After Doubling](https://startupfortune.com/openais-revenue-run-rate-tops-40-billion-just-months-after-doubling/)", "url": "https://wpnews.pro/news/china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests", "canonical_source": "https://startupfortune.com/chinas-zai-says-its-new-model-nears-anthropics-mythos-5-on-cyber-tests/", "published_at": "2026-08-14 10:50:34+00:00", "updated_at": "2026-08-14 20:49:51.872068+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-policy", "ai-research", "ai-products"], "entities": ["Z.ai", "Zhipu AI", "GLM-5.2", "GLM-5.3", "Anthropic", "Mythos 5", "Project Glasswing", "Amazon Web Services"], "alternates": {"html": "https://wpnews.pro/news/china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests", "markdown": "https://wpnews.pro/news/china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests.md", "text": "https://wpnews.pro/news/china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests.txt", "jsonld": "https://wpnews.pro/news/china-s-z-ai-says-its-new-model-nears-anthropic-s-mythos-5-on-cyber-tests.jsonld"}}