China’s Zhipu AI (Z.ai) claims its open-weight GLM-5.2 model can match Anthropic’s Mythos on cybersecurity vulnerability detection, even if it still lags behind on general reasoning tasks.
The claim, reported by The Verge, marks a significant new frontier in the AI model race. GLM-5.2 was originally released in June as a coding-focused model with a 1-million-token context window and an MIT license, making it freely available for self-hosting, fine-tuning, and commercial use anywhere in the world. Independent benchmarks showed it rivaling GPT-5.5 on coding tasks at roughly one-sixth the cost.
The cybersecurity angle is newer. Researchers testing GLM-5.2 against Mythos on bug-finding benchmarks found that the gap between the two models has narrowed substantially on vulnerability detection, even though Mythos remains stronger on general-purpose reasoning. Given GLM-5.2’s MIT license and cost advantage, roughly US$1.40 (approximately £1.10) per million input tokens versus Mythos’ premium pricing, the model could lower the barrier for mid-sized companies that previously found advanced AI security tools too expensive.
Z.ai is one of several Chinese AI labs that have stepped into the gap left by US export controls. Anthropic’s Mythos 5 was briefly disabled for foreign nationals in June after a US government order, creating demand for models outside American jurisdiction. Z.ai and others have positioned their open-weight releases as sovereign AI alternatives that cannot be switched off by Washington.
Independent benchmarking is still needed to verify Z.ai’s cybersecurity claims. Early reports should be treated as preliminary until standardized testing confirms the performance against Mythos at scale.
Sources: China’s Z.ai claims it can match Mythos on cybersecurity (The Verge, June 28, 2026); Zhipu AI releases GLM-5.2 (DataNorth, June 18, 2026)