At a time when Anthropic is throwing out the “AI deceleration theory,” the commercialization of AI is speeding up! From model development to financial advisors, accelerate the redemption of AI agent dividends

Zhitong Finance App learned that the CEO of Anthropic, the world's strongest leader in AI applications, just said “the development of AI models/AI technology should slow down” last weekend. Immediately after that, there was news that the next-generation divine model Claude Opus 5.2 had already started grayscale testing in Claude Code, and the latest news also showed that the company is making every effort to accelerate the penetration of Anthropic's Claude series of AI application tools into various industries — there is news that Anthropic launched a major launch Claude's new AI tool for financial/financial advisors aimed at Wall Street financial giants such as BlackRock. Together, these developments can be described as highlighting that Anthropic is simultaneously promoting cutting-edge AI risk management and commercialization of enterprise AI applications, and competing for AI monetization with the strongest AI application competitors such as OpenAI.

Many AI application developers in the AI open source ecosystem have discovered that the new generation of “god-level models”, Claude Opus 5.2, has already started grayscale testing in Claude Code. Furthermore, some Claude ecosystem developers discovered by capturing packages and checking the request status (/status) that although the front-end name has not changed, the model behind it, Slug, clearly points to Opus 5.2.

Anthropic aims to advance AI from programming aids to work systems that can continuously complete complex tasks through stronger model capabilities and intelligent execution mechanisms. In the latest descriptions of Opus 5.2 from developers around the world, notable upgrade directions include responsiveness, code completion, and continuous execution and repeated verification capabilities during long tasks. Among them, the most direct change in commercial value is that after users deliver a goal, AI can take on more tasks of teardown, code writing, testing, and repair, reducing repeated manual prompts and takeover.

Regarding Opus 5.2, some developers even said that it seems that recursive self-improvement (RSI) is beginning to dominate the Anthropic training paradigm, showing a technical path of “stronger models to assist R&D, improve R&D efficiency, and drive the next generation of models.” The potential significance of RSI is that AI models tend to be automated product and AI R&D tools at the same time, extending competition among model companies to AI laboratory training execution work, infrastructure maintenance, and automation of cutting-edge theoretical research processes for underlying operators.

On September 14, the stock price of Nvidia, the “AI chip superpower,” fell by about 3.4%, while the Philadelphia Semiconductor Index rarely fell sharply by about 6%. The market is taking into account the risks brought about by factors such as AI deceleration discussions — global AI leaders such as Anthropic and OpenAI unanimously called for a slowdown in the development of cutting-edge AI models over the weekend.

However, Anthropic, which introduced the “AI deceleration theory,” can be described as making every effort to accelerate the AI application-side monetization path. In addition to the Opus 5.2 grayscale test, Anthropic also announced the major launch of the Claude AI Financial Advisor Edition on Monday, connecting data and analysis tools from top Wall Street financial institutions such as BlackRock, Pioneer Pilot, and Carson Wealth Management to help advisors prepare customer meetings, review portfolios, sort out records, and draft communication materials. On the one hand, the company is calling for a slowdown in the development of cutting-edge model capabilities, but on the other hand, it is actively promoting the development of cutting-edge AI models and promoting the integration of the company's AI models into existing business processes to seek revenue generation data for larger enterprise customers and AI applications.

Frontier AI developers are calling for a slowdown, but the application side is racing to monetize

Grayscale testing of the new AI model and the release of the new Claude AI application tool are positive proof that while Anthropic is calling for a slowdown in the AI R&D process, it continues to accelerate the commercialization of AI applications. Judging from the active AI application product layout of the world's leading AI application manufacturers, AI tools based on cutting-edge AI models are entering business scenarios with specific processes such as finance, content creation, medical care, and scientific research.

OpenAI launched ChatGPT for the financial services industry on September 10, designed in collaboration with Morgan Stanley and Evercore to comprehensively and deeply integrate the GPT-6 Astra model, professional financial data and document generation capabilities, first serving investment banking and stock research; Roblox expanded the AI game creation tool Build on September 11 and announced independent application and browser gameplay plans; Anthropic launched tools for medical institutions and expanded life science functions at the beginning of this year to support insurance pre-authorization data Processing, scientific research information retrieval, and preparation of regulatory application materials.

These latest trends in AI applications all show that competition is extending to who can embed models into high-frequency, verifiable workflows, and some people are willing to pay. However, product launch, pilot use, and full commercial deployment are at different stages. Currently, the number of announcements cannot be directly used as the industry penetration rate.

Recursive self-improvement has undoubtedly become an important research direction for public discussion at the world's most advanced AI laboratories. Anthropic recently revealed that its quarterly code delivery volume per engineer has reached about 8 times the level of 2021-2025. At the same time, it clearly stated that the closed loop of being able to completely independently design and develop next-generation models has not yet been achieved, and there is still a clear gap in research target selection and judgment capabilities. OpenAI's chief scientist also publicly stated that the company is shifting its research focus to RSI. The coverage of AI-assisted R&D work is increasing significantly, and it is possible to shorten most AI R&D work processes through code generation, experiment execution, and result analysis.

The main impact of Anthropic's cutting-edge AI model, which is undergoing grayscale testing, and cutting-edge models such as OpenAI Astra, which is currently undergoing grayscale testing, on computing power requirements focuses on making more complex tasks executable, thereby expanding the scope of potential use. More paid agent tasks are expected to increase inference requirements over the long term.

The continuous expansion of enterprise applications can be described as an important source of continuous acceleration in computing power demand. A financial advisory task may include reading positions, searching for research, operation analysis, checking results, and generating customer materials, requiring multiple model calls and external tool execution; if more customers and institutions hand over such tasks to agents, the cumulative amount of reasoning, concurrent sessions, and tool operation resources may continue to increase dramatically; long contexts, multi-step reasoning, and parallel sub-agents may also increase resource requirements for complex tasks.

In the AI data center computing power infrastructure chain, the demand for strong computing power resources brought by AI agents may spread rapidly to various links such as GPU/ASIC, HBM, server DRAM, enterprise-grade SSD, high-speed optical interconnection equipment within data centers, data center CPUs, data center power chains, etc. Agent expansion affects computing, memory, storage, and networking simultaneously. GPUs and other accelerators are responsible for model calculation, and high-bandwidth memory (HBM) supports high-speed access to model weights and active inference states; long contexts and concurrent sessions increase key value cache (KV Cache) pressure, while CPU-side DDR memory undertakes tasks such as tool execution, database access, and session management. Enterprise-grade NAND solid-state drives (SSDs) are used for knowledge bases, documents, and operation records, and can handle partial KV cache offload and reuse under an appropriate architecture. Micron Technology's recent technical note also divided these requirements into levels such as high-speed memory close to accelerators, data center main memory, and context storage.

The world's top wealth management institutions are starting to install AI assistants: Claude struggles to win the Wall Street Financial Advisor Workbench

According to information, Anthropic is introducing a new version of Claude to financial advisors at Wall Street's top asset management institutions and comprehensive financial institutions, combining this chatbot with financial analysis and risk management technology provided by BlackRock, Pioneer Pilot Group, and others.

According to senior management of Anthropic and BlackRock, this AI intelligent operating system called “Claude for Financial Advisors” (Claude for Financial Advisors) is expected to speed up the processing of tasks such as research, administrative affairs, and portfolio supervision.

This is one of the most important steps the AI company has taken to date to expand into the financial industry. This feature can also be connected to the tools of companies such as Carson Wealth Management and iCapital, and is further developed on the basis of financial service AI agents previously launched by Anthropic; these agents are designed to handle financial service tasks such as producing business promotion presentations and reviewing reports.

As Anthropic introduces products to the financial industry, the entire AI industry is in a period of both growth and turmoil. Anthropic and OpenAI are both planning initial public offerings that will generate billions of dollars in revenue for early investors. OpenAI just launched financial services features tailored to investment bankers and stock researchers last week.

Meanwhile, the rapid development of artificial intelligence technology is alerting legislators and industry leaders around the world. On Saturday, Anthropic CEO Dario Amoudi said development of the most advanced systems must be slowed down to prevent a disaster; OpenAI CEO Sam Altman and SpaceXAI CEO Elon Musk both supported his statement.

Like OpenAI, Anthropic has been vying for enterprise customers through professional services outside of software engineering and programming.

Claude Financial Advisor Edition is part of this initiative, which focuses on providing advisors with more efficient workflows so they can serve more customers.

Jonathan Pelosi, head of financial services at Anthropic, said in an interview: “The people who actually work as financial advisors — aren't that big, and they're actually shrinking. These people are retiring, and this group is already small in size. As a result, the supply of high-quality financial guidance is actually inadequate. If there's anything we can do to help these advisors serve more clients, we think that's definitely a good thing.”

Anthropic's tools allow more advisors to access portfolio analysis and investment research tools from companies such as BlackRock and Pioneer Pilot, and may bring more business to these companies. Financial advisors are increasingly relying on “model portfolios” made up of exchange-traded funds and other investments; BlackRock, for example, said it currently manages about $300 billion of such portfolios.

Jamie Magiella, BlackRock's US Wealth Advisor and Head of Retirement Business, said, “One of the biggest trends we've observed is that advisors want to outsource their work. The opportunity to help advisors build their portfolios is not only about providing better information, but also about helping them improve their service capabilities.”

Investment advice and investment decisions will remain the responsibility of advisors and their clients.

Pelosi said, “You don't get investment advice directly from Claude. We leave that judgment to the professionals.”