GLM 5.2 Model on OpenRouter Shows Significant Discounted Pricing
开发生态
According to the OpenRouter platform, some providers are currently offering API services for GLM 5.2 with discounts as low as 0.5% of the original price. The specific reason and duration for this discount have not yet been clarified.
According to the OpenRouter platform, the API pricing for Zhipu GLM 5.2 is currently heavily discounted, with some inference service providers cutting prices by 95%. Input prices are as low as $0.07 per million tokens, and output prices are $0.22 per million tokens.
Report: Hassabis Originally Wanted to Leave but Was Persuaded to Become DeepMind Chairman
行业动态
According to reports, former Google DeepMind CEO Demis Hassabis originally wanted to step down at the same time as Jeff Dean, but Google management, concerned that the news could cause a stock price plunge, persuaded him to serve as chairman temporarily until the situation stabilizes before he exits gracefully. This claim has not been officially confirmed.
According to Pathfounders, citing industry sources, former Google DeepMind CEO Demis Hassabis originally wanted to leave Google at the same time as Jeff Dean, but Google management, concerned that the news would cause the stock price to plummet, convinced him to temporarily serve as chairman of Google DeepMind before exiting gracefully once the situation stabilizes. This claim has not been officially confirmed. Pathfounders also predicts that Hassabis will leave Google within a year.
Envision Ulanqab Xinghe Base Commences Production, with Planned Capacity Reaching 2GW
行业动态
It is reported that the 'Envision Ulanqab Xinghe Base' has officially commenced production. According to officials, the base has a total planned capacity of 2GW, with million-card parallel computing capability and million-P computing scale, making it the world's largest single AI data center in terms of token output capacity and the largest AI computing park in the country.
According to reports, Envision Energy Group announced the completion of an AI computing super-monolith in Ulanqab, with the 'Envision Ulanqab Xinghe Base' officially put into operation. Envision officials said the super-monolith has a building volume of 120,000 square meters, with million-card parallel capability and a million-P computing scale, making it the world's strongest single AI data center in token output capacity and the country's largest AI computing campus. The base's total planned capacity is 2GW, featuring a very high proportion of direct green power connection, with green power supplied directly through self-built wind farms and dedicated lines. This is the first flagship project of Envision's 'Gobi Mission', aimed at creating a replicable solution for domestic computing clusters.
Claude Code Lead Says Claude Has Largely Solved Prompt Injection Threats Through Multi-Layer Defense
技术与洞察
Claude Code lead Boris Cherny posted that Anthropic, through multi-layer defense involving models, detectors, and intent classifiers, has largely solved prompt injection threats, reducing the success rate of unseen attacks to nearly zero.
Boris Cherny, the lead creator of Claude Code, posted that prompt injection is the most common way to attack users and agents, and early Claude models were affected by it. He pointed out that Anthropic has achieved unexpectedly positive results by training models to defend against such attacks, essentially solving this threat in actual use of Claude models. Boris further explained that by stacking multiple layers of defense such as models, input detectors, and intent classifiers, the success rate of unseen prompt injection attacks can be reduced to near zero. Additionally, he believes it is safer to use subagents that cannot access credentials like SSH passwords to read untrusted sources.
Note: Content is AI-assisted and may contain hallucinations and errors.
