Anthropic is simultaneously advancing frontier AI risk governance and the commercialization of enterprise AI applications, competing with OpenAI and other rivals in AI monetization. On the one hand, it calls for slowing down the development of frontier models; on the other hand, it actively pushes forward model research and development and integrates them into existing business processes, striving for larger-scale enterprise customers and AI application revenue data. Product launches, pilot use, and full commercial deployment belong to different stages, and the number of announcements cannot currently be directly equated with industry penetration rates. What is certain, however, is that AI tools based on the most frontier AI large models are entering business scenarios with specific processes in finance, content creation, healthcare, and scientific research, and computing power demand will continue to grow accordingly.
Dario Amodei, CEO of Anthropic, the global leader in AI applications, publicly stated last weekend that the development speed of the most advanced AI systems must be slowed down to prevent catastrophe. OpenAI CEO Sam Altman and SpaceXAI CEO Elon Musk both expressed support for this. However, Anthropic has recently been accelerating the commercialization process on the AI application side, forming a sharp contrast.
The new generation “god-tier model” Claude Opus 5.2 has begun gray-scale testing in Claude Code. Developers discovered through packet capture and checking request status that although the front-end name remains unchanged, the back-end model slug already points to Opus 5.2. According to reports, this model has upgrades in response speed, code completion quality, and sustained execution and repeated verification capabilities in long tasks. After a user delivers a goal, the AI can take on more task decomposition, code writing, testing, and repair work, reducing manual repeated prompting and intervention.
Some developers say that recursive self-improvement (RSI) seems to be beginning to dominate Anthropic’s training paradigm, demonstrating a technical path of “stronger models assisting research and development, improved R&D efficiency, and driving the progress of next-generation models.” Anthropic recently disclosed that its engineers’ per capita quarterly code delivery volume reached approximately 8 times the level from 2021 to 2025, while clearly stating that a closed loop capable of fully autonomously designing and developing next-generation models has not yet been achieved. OpenAI’s chief scientist also publicly stated that the company is shifting its research focus to RSI.
Anthropic announced on the same day the launch of “Claude Financial Advisor Edition,” connecting data and analytical tools from top Wall Street financial institutions such as BlackRock, Vanguard, and Charles Schwab to assist advisors in preparing for client meetings, reviewing investment portfolios, organizing records, and drafting communication materials. Jonathan Pelosi, head of financial services at Anthropic, said that the number of people truly working as financial advisors is small and shrinking, and the supply of high-quality financial guidance is not abundant, so helping these advisors serve more clients is a good thing. Jamie Maggiela, head of U.S. wealth advisory and retirement at BlackRock, said that advisors want to outsource work, and the opportunity to help advisors build portfolios lies not only in providing better information but also in improving service capacity. Investment advice and investment decisions remain the responsibility of advisors and their clients.
OpenAI launched ChatGPT for the financial services industry on September 10, designed in collaboration with Morgan Stanley and Evercore, combining the GPT-6 Astra large model, professional financial data, and document generation capabilities. Roblox expanded its AI game creation tool Build on September 11. Anthropic launched tools for healthcare institutions earlier this year and expanded life sciences features. These developments show that competition is extending to who can embed models into high-frequency, verifiable workflows that people are willing to pay for.
