Z.ai released GLM-5.3 without retraining the base model, scoring 84.5% on the CyberGym cybersecurity benchmark (ahead of Claude Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%) with a 50% coding improvement on Terminal Bench 3.0 and Agents' Last Exam. The model found 1,097 critical bugs through post-training exploit chains, and open-source weights are promised within two weeks.
Independent AI benchmarking startup Vals AI closed a $40M Series A at a $400M valuation led by Andreessen Horowitz, after revealing that frontier models fail 52% of real-world finance analyst tasks. The funding will expand independent benchmarks across coding, cybersecurity, and professional domains.
Qualcomm's Snapdragon X2 Elite Extreme won every CPU and AI benchmark against Intel Core Ultra X9, even on battery power in Balanced mode, and surpassed AMD Ryzen AI 9 465 by 83% in Geekbench 7 according to Signal65 testing. The results suggest Qualcomm's latest platform combines high compute throughput with aggressive power efficiency for premium AI laptops.
Nvidia set up a financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize capital for AI infrastructure, shifting its competitive moat from chip design to financial leverage. Dell'Oro Group projects data center capex will surpass $3 trillion by 2030 as the AI buildout maintains momentum despite chip shortage concerns.
New LLM benchmark comparisons show open-weight models closing the gap with proprietary leaders, with GLM 5.3 scoring 94 and Qwen 3.8 Max jumping from 51 to 92 on Ruby coding tasks. Netlify now runs any OpenRouter model in production including Kimi K3, GLM 5.2, and DeepSeek V4, testing 11 models on the same build prompt to compare real-world performance.
GLM-5.3's post-training process spontaneously produced exploit chains that Z.ai did not design or anticipate, identifying 1,097 critical vulnerabilities in cybersecurity testing. On the harder ExploitBench and ExploitGym benchmarks, GLM-5.3 still trails Claude Fable 5 (which reaches 39.5% at maximum effort), but the open-weight model's autonomous discovery capability surprised even its developers.