Guotai Haitong: Anthropic launches new model, expecting to promote the development of sovereign AI layout.

Guotai Haitong released a research report stating that Anthropic released Claude Fable 5, which surpasses all previously publicly released models in almost all benchmark tests in software engineering, knowledge work, vision, and scientific research. It is the strongest AI model for humans to date. At the same time, it launched the cutting-edge model Mythos 5 for security defense, with anti-steaming security restrictions, offering a free trial window from June 10th to June 22nd, priced at $10 per million input tokens, and $50 per output token. Subsequently, due to the impact of US export controls and security requirements, both models were globally taken offline, directly hindering the overseas use of advanced models, short-term driving the demand for domestic large models to return, and in the medium to long term will further drive the demand growth for local computing power, self-developed models, and the privatization and sovereignty of AI models. Guotai Haitong's main points are as follows: Anthropic released Claude Fable 5, which is considered the most powerful model ever, leading in almost all AI performance tests. In software engineering, Fable 5 compressed a 50 million lines of Ruby codebase in one day, compared to over two months manually. In the Cognition Frontier Code assessment, it scored the highest with moderate effort. In knowledge work, Fable 5 scored the highest in the Hebbia Finance Benchmark among all models. It has become the new SOTA for visual tasks, able to extract precise numerical values from scientific graphs and even reconstruct Web application source code from screenshots. Fable 5 can handle millions of tokens in long-term tasks, with performance in "Slay the Spire" performance 3 times higher than Opus 4.8. Claude Mythos 5 was also introduced, targeting network security defenders and infrastructure providers through Project Glasswing deployment, demonstrating remarkable capabilities in scientific hypothesis formulation and drug design fields. With Mythos 5, Anthropic increased the efficiency of certain stages of the drug design process by about 10 times. Mythos 5 is the first model to continuously propose novel scientific hypotheses: compared to blind tests at Opus level, about 80% of scientists prefer Mythos' hypotheses. The pricing is $10 per million input tokens and $50 per output token, less than half of Mythos Preview, but twice as expensive as Opus 4.8. Subscribers can use Fable 5 for free from now until June 22nd, and the free access will be removed on the 23rd. Just a few days after the release, Fable 5 was subject to US government export controls, limiting global user access or accelerating demand for sovereign models. Just days after the model was released, the US government, citing national security concerns, instructed Anthropic to suspend all access to Fable 5/Mythos 5 for "foreign nationals" (including foreign nationals and employees of foreign companies) in the US; as the company found it difficult to accurately distinguish by nationality in real-time, it decided to implement a "global full shutdown" of the two models. The official reason given was a narrow-range jailbreak method that could bypass security barriers and allow the models to access underlying capabilities in cybersecurity scenarios. Anthropic disagreed but complied with the regulations. In the short term, the industry believes that domestic users' reliance on Fable 5 will be hindered, accelerating the return of demand to local models and private deployment; in the medium to long term, investment shift from "renting intelligence" to "building sovereignty" will drive the demand for local computing power and models. Risk Warning: Technological developments may not meet expectations, and market competition may increase risks.