Recent developments in the AI landscape highlight significant shifts in API pricing, model performance, and internal research. DeepSeek has transitioned from flat-rate pricing to a time-of-day model effective August 17, introducing peak and off-peak rates to manage capacity. Under this new structure, V4-Pro output costs range from ¥27 ($3.96) per million tokens during peak hours to ¥13.5 ($1.98) during off-peak windows, while V4-Flash scales from ¥9 to ¥4.5. Simultaneously, Zhipu has released GLM-5.3, an open-weights coding model that utilizes post-training scaling to achieve significant benchmark gains. GLM-5.3 reached a score of 28.3 on Terminal-Bench 3.0 and 66.9 on DeepSWE v1.1, while its CyberGym score of 84.5% rivals Claude Mythos 5. Notably, the model identified 2,404 vulnerabilities during testing, including a long-standing DNS protocol flaw. Finally, Anthropic’s latest risk report revealed "Model 2," an internal-only model surpassing Claude Mythos 5 with a 62.8% CoBench v2 score compared to the public version's 50.3%. Despite this capability, Anthropic has no plans to release Model 2. The report also admitted a biosafety classifier was inactive for approximately 11 months, allowing roughly 133 million conversations involving 50,000 contractors to bypass safety protocols.