Full Summary
This Friday, Chinese AI firm Z.ai has unveiled its new GLM-5.2 open-weight model, with multiple sources including LatestLY, Memeburn, and Storyboard18 confirming its launch. This 753-billion-parameter model boasts a one-million token context window and is available under an MIT license, making its core weights publicly accessible. Here's the thing: Z.ai claims GLM-5.2 outperforms OpenAI's GPT-5.5 on coding benchmarks and "Humanity’s Last Exam," a finding supported by Storyboard18. Memeburn adds that it costs about one-sixth the price of competing models. Z.ai's founder, Tang Jie, suggests China could develop a "Mythos-class" AI model, similar to the powerful Mythos 5 and Fable 5, before 2027, as reported by India Today and News9live. This comes after the US restricted foreign access to those advanced models. What nobody expected is that despite these new advancements, the AI benchmark AA-Briefcase, announced by Artificial Analysis, shows even top models like Claude Fable 5 only satisfy all rubric criteria on three percent of knowledge work tasks. Blockchain.news reports Claude Fable 5 leads with an Elo score of 1587, but costs vary wildly from over $31 to just four cents per task. Meanwhile, NVIDIA has released Nemotron 3 Ultra, an open-weight model designed for long-running AI agents, focusing on multi-step processes like planning and code execution. And in a completely different area, IBM watsonx.ai 2.4 expands access to governed models, including a new bundle with Red Hat AI Enterprise. This means that while AI models are rapidly advancing, offering more affordable and powerful options for coding and specialized tasks, their real-world application in complex knowledge work remains a significant challenge, impacting businesses' ability to fully automate high-stakes projects.