According to Coinbase.com, CEO Brian Armstrong stated that Coinbase is evolving into a full-fledged exchange and will fully advance its AI Financial Account, allowing users to assign wallets to AI agents, receive AI-driven financial advice, and connect Coinbase accounts to major mainstream models. Coinbase will restructure Coinbase Advanced, integrating global liquidity between US and international users and between Coinbase and Deribit. The Coinbase Developer Platform will provide enterprises with stablecoin payment capabilities and a compliant escrow account system, and will launch a developer tools dashboard. Base has introduced private trading functionality and launched a version of the Base web application.

According to a report by CryptoWorld, OpenAI has published new research confirming a remarkable "strong generalization" phenomenon in the field of alignment: researchers only need to teach AI to be a "good person" who doesn't lie and is willing to admit mistakes in a very limited number of everyday scenarios, and the AI will automatically behave well in completely new scenarios it has never learned before. Experiments have shown that positive alignment traits also have cross-domain contagion power. The research team focused on training large models in the medical, educational, and scientific fields to cultivate traits such as honesty and humility. Tests revealed that after learning the traits of good people, the models outperformed ordinary models in 44 out of 53 security assessments. The research team believes that injecting reinforcement learning training with kindness allows alignment traits to take root deep within the model, truly achieving a "point-to-area" protective effect.

According to CoinJie.com, Coinbase CEO Brian Armstrong stated on the X platform that Coinbase now includes pre-IPO perpetual contracts and stock options, and will soon support tokenized stocks. In addition, Coinbase has redesigned Coinbase Advanced and begun integrating global liquidity between US and international users, and between Coinbase and Deribit users. CoinbaseDev is providing stablecoin payment capabilities for businesses, launching fully custodied accounts based on its compliant technology stack, and releasing a new developer tools dashboard. Regarding Base, Coinbase announced the launch of private trading and a web-based Base app. Coinbase is also becoming a financial account for AI, supporting wallets for AI agents, access to AI-driven financial advice, and connecting Coinbase accounts to commonly used LLMs.

According to Coinbase.com, Coinbase CEO Brian Armstrong stated at an event that this week has been significant for Coinbase, thanking all Coinbase employees and their thousands of AI assistants for their hard work and dedication that made all the announcements possible. Coinbase's all-in-one exchange now includes pre-IPO perpetual contracts, stock options, and upcoming tokenized stocks. Furthermore, Coinbase has redesigned its premium version and begun integrating global liquidity, bringing the benefits of stablecoin payments, supporting businesses with fully custodied accounts, and launching a brand-new developer tools dashboard. Coinbase also announced private trading and basic applications on its website. Finally, Coinbase is becoming an AI-powered financial account, allowing users to provide wallets to their agents, receive AI-driven financial advice, and connect their Coinbase accounts to their favorite LLMs.

According to CoinWorld, AA-Briefcase reported that Claude Fable 5 topped the benchmark, while GLM-5.2 made it into the top three. Artificial Analysis, the benchmarking organization, launched its first long-cycle knowledge work benchmark designed specifically for large-scale intelligent agents, covering four scenarios: data science, product management, banking operations, and heavy industry strategy. Developed by industry experts from Google, McKinsey, and Boston Consulting Group, it includes 91 tasks designed to simulate complex real-world business project flows. The results show that Claude Fable 5 achieved the highest overall score, with Claude Opus 4.8 and GLM-5.2 ranking second and third, respectively. Despite Claude Fable 5's strong performance, its perfection rate was only 3% under the strict single-task perfect score standard. In the open-source model category, GLM-5.2 performed exceptionally well, with an overall score only 90 points lower than Claude Opus 4.8, but with less than 25% of the operating cost.