Anthropic opens Mythos level model Claude Fable 5 to the public, introducing new security measures and paving the way for a potential IPO.

Notice that, citing the risk of falling into the hands of criminals, Anthropic has only opened its Mythos model to a limited number of users for two months, the company announced on Tuesday that it is preparing to release a new, equally powerful model to the public. Anthropic has introduced Claude Fable 5, a model belonging to the Mythos level that will be available to enterprise customers and paid subscribers. The company stated that the reason for the large-scale release is the introduction of new security measures (Safeguards) that can intercept violations in specific high-risk areas such as network security and biotechnology. Diane Payne, the head of product management in Anthropic's research department, said in an interview, "For us, this is about what we call 'striving upstream' - being able to provide this technology in a valuable way, while ensuring that its benefits far outweigh the harm through appropriate protective barriers." In April of this year, Anthropic stunned Wall Street and government officials with its outstanding performance in identifying software vulnerabilities using the Mythos model. At that time, the company stated that it did not intend to widely promote the model and would restrict it to a few selected companies in a network security plan called "Project Glasswing." However, with the release of Claude Fable 5, Anthropic has fulfilled its "ultimate goal" of deploying Mythos-level models on a large scale. At the same time, this move leverages the company's growing technological momentum and investor interest on the eve of the anticipated massive IPO later this year. Anthropic claims that Claude Fable 5 has shown "excellent performance" in software engineering knowledge and knowledge work tasks. According to an official blog post, in certain benchmark tests, it scored more than 10% higher than another model released by the company last month, Claude Opus 4.8. Payne pointed out that Claude Fable 5 represents a "significant leap forward" in capabilities, which is why the company must implement additional protective barriers to prevent misuse. For example, if a user asks high-risk questions like how to make toxins such as ricin, the model will intercept the response and revert to Claude Opus 4.8 to provide a safe answer. "We hope to consciously build new classifiers and security mechanisms in this release," Payne said. On the same day, Anthropic also launched an updated Mythos model called Claude Mythos 5. According to the blog, its underlying model is the same as Claude Fable 5, but removes some security restrictions in certain areas. Just a few days ago, Anthropic secretly submitted its IPO prospectus to the U.S. Securities and Exchange Commission (SEC). After experiencing explosive growth this year, the company is preparing for a potentially historic stock offering. In May, Anthropic stated that its annual recurring revenue had surged to $47 billion, far exceeding last year's approximately $10 billion. The company's recent financing valuation reached $965 billion, surpassing its main competitor OpenAI, which was valued at $852 billion at the end of March. OpenAI is also preparing for a major IPO and announced on Monday that it had secretly submitted its prospectus to regulatory agencies. Elon Musk's SpaceX is scheduled for a record-breaking public market debut on Friday. Intense competition means that Anthropic CEO Dario Amodei will face pressure to prove the company's high valuation to investors, and Claude Fable 5 could become a valuable new revenue growth point for the company. The model costs $10 per million input tokens and $50 per million output tokens, which is twice the price of Claude Opus 4.8. Payne stated that pricing is indeed the "primary consideration" for customers, but they are not just looking for low cost. She said that users are seeking higher accuracy and return on investment, and early customers of Claude Fable 5 have noticed optimizations in the cost of each task. "By using more intelligent models, you can achieve a higher return on investment," Payne concluded.