China Securities Co., Ltd.: Mythos triggers the expansion of demand for cutting-edge model training, and the Google TPU industrial chain is expected to continue to benefit.

China Securities Co., Ltd. issued a research report stating that Mythos is the strongest model to date according to Anthropic, with its capabilities experiencing a leap-like enhancement, leading in code, reasoning, and research assistance comprehensively over its predecessors. Its most outstanding advantage is seen in real-world network security scenarios, surpassing Opus 4.6 by far in tasks such as Firefox vulnerability exploitation, and it can also discover and exploit zero-day vulnerabilities. Additionally, Mythos demonstrates a significant acceleration effect in the research and development processes of industries such as biochemical research, finance, AI training, and Siasun Robot & Automation. Anthropic is backed by investments and computing power from Amazon and Google, and has recently strengthened cooperation. Starting from 2027, the demand for TPU is expected to reach 3.5GW, and the Google TPU industry chain is expected to continue benefiting from the expansion of advanced model training demands. The main points of China Securities Co., Ltd. are as follows: 1. The strongest model to date, with capabilities experiencing a leap-like enhancement. Claude Mythos is defined by Anthropic as the "strongest AI model to date," with a clear leap in its capabilities. In various dimensions such as software engineering, reasoning, knowledge work, and research assistance, it is significantly stronger than Opus 4.6. For example, in terms of coding ability, Mythos reaches 93.9% on SWE-bench Verified, significantly higher than Opus 4.6's 80.8%; in terms of pricing, its API pricing is $25 per million input tokens and $125 per million output tokens, roughly five times that of Opus 4.6. 2. Extremely strong Cyber capabilities, particularly displaying a significant lead in real-world scenarios. In tasks like Firefox vulnerability exploitation more closely related to the real world, Mythos achieves an overall success rate of 84.0%, with a complete vulnerability exploitation success rate of 72.4%, far exceeding Opus 4.6's 15.2%. Additionally, Mythos can discover and exploit zero-day vulnerabilities in real systems, exemplified by uncovering a 27-year-old OpenBSD vulnerability, indicating that its cyber capabilities have transitioned from "being able to solve secure problems" to "being able to find and exploit vulnerabilities in a real environment." 3. It has a significant amplification effect in biochemical research and various industry research and development processes. Besides Cyber, the value of Mythos is also beginning to manifest clearly in chemistry, biology, and broader research scenarios. In the biochemical field, its integration capabilities for public documents, experiment processes, multi-step protocols, and biological sequence design/functional prediction have significantly improved; on a broader industrial front, it is noticeably stronger than its predecessors in tasks such as Kernel task, LLM training, with some tasks reaching the equivalent of several hours to 40 hours of human work, and is expected to accelerate the research and development processes in multiple industries such as biomedicine, financial quantification, AI infrastructure, Siasun Robot & Automation, and autonomous driving. 4. Current alignment capabilities are the best, but the risk of infrequent errors is higher. Mythos is described as the model with the "best overall alignment" to date. After training, the model's cooperation rates in destructive behavior, cheating behavior, and misuse requests towards humans have continued to decrease compared to its predecessors. However, due to the rapid advancement of its capabilities, in very few cases, if it continues execution along an incorrect path, its ability to covertly achieve additional goals and evade monitoring is stronger. 5. Anthropic: Financing and computing power cooperation concurrently intensified, with Google and Amazon as core partners. Anthropic's valuation as of April 2026 has exceeded $380 billion, with an ARR surpassing $30 billion. Amazon has invested a cumulative $8 billion and has deeply integrated Anthropic's core training through AWS and Trainium; Google has invested over $30 billion, starting cooperation in October 2025, and in April 2026, Anthropic further increased the deployment volume of TPU for the next year, with an expected future capacity of 3.5GW. Based on power consumption estimates, it is predicted to involve the purchase of 200-300 million chips. 6. Google TPU: Dual iteration of microarchitecture and cluster system, expected to continuously benefit from the expansion of advanced model training demands. The core advantage of Google TPU lies not only in the performance improvement of a single chip but in the "chip + interconnect + cluster" integrated evolution. From TPUv1 to TPUv7, Google has continuously upgraded in matrix computing units, HBM, ICI interconnection, topological structure, and cluster capabilities, transforming TPU from an early inference chip to a complete supercomputing platform capable of supporting large model training. With the deepening cooperation between Google and leading AI companies on the next-generation TPU cluster, TPU is becoming an important foundation for advanced model training. Risk Warning: Development of artificial intelligence technology may not meet expectations; capital expenditures of Internet companies may fall short of expectations; competition intensifies due to the involvement of numerous companies; policy and regulatory efforts may not meet expectations.