How Anthropic Turned a Restricted AI Model Into an Enterprise Product

Anthropic spent the past two months limiting access to one of its most advanced AI models. In April, the company introduced Mythos Preview through Project Glasswing , a program that gave governments, cybersecurity firms, infrastructure operators, and select enterprises access to a model Anthropic believed possessed unusually advanced cybersecurity capabilities. The company initially restricted access because of concerns about how those capabilities could be used if released broadly. On June 1, Anthropic confidentially filed draft paperwork for a proposed IPO . Eight days later, it launched Claude Fable 5, a public version of the Mythos model family that includes safeguards designed to limit some of the underlying system's most sensitive capabilities. The release arrives as AI companies increasingly focus on enterprise systems that can perform work across business processes rather than function solely as chat assistants. AI investment is increasingly moving beyond standalone software tools and into operational systems . From Project Glasswing to Fable 5 Anthropic's path to Fable 5 differed from a typical AI model launch. Rather than immediately releasing Mythos to developers and enterprises, the company created Project Glasswing , a controlled-access program focused on cybersecurity and infrastructure protection. Participants included cybersecurity organizations, critical infrastructure operators, government agencies, and selected enterprises. Anthropic said the goal was to evaluate how advanced AI systems could help identify and address software vulnerabilities while reducing the risk of misuse. The program produced results that went beyond benchmark testing. In a recent update , Anthropic said Mythos Preview identified 23,019 potential software vulnerabilities across open-source projects, including an estimated 6,202 high- or critical-severity vulnerabilities. The company said organizations participating in Glasswing were also using the model for patch generation, pre-release security reviews, penetration testing, threat detection, and legacy code modernization. Anthropic gradually expanded access to the program during the spring. Reports emerged that unauthorized users had accessed Mythos, highlighting the challenges associated with controlling access to increasingly capable AI systems. Fable 5 represents Anthropic's attempt to bring that technology to a broader audience without fully releasing Mythos itself. According to Reuters , certain cybersecurity and biosecurity requests are redirected away from Fable and handled by Claude Opus 4.8. Anthropic says the approach allows enterprises to access much of the system's capabilities while maintaining restrictions around higher-risk use cases. Enterprise Software Companies Are Building for AI Agents Anthropic's launch comes as enterprise software vendors increasingly focus on AI systems capable of executing tasks across multiple applications and workflows. At Snowflake Summit 26, Snowflake centered its product strategy around what it called the "Agentic Enterprise." The company introduced a range of capabilities designed to help organizations deploy AI agents across business processes, data environments, and operational workflows. The announcement is part of a broader movement among infrastructure vendors to build enterprise systems around AI agents rather than standalone assistants. OpenAI has followed a similar path. Over the past year, the company introduced the Responses API and Agents SDK, tools designed to help developers build systems that can use software tools, access information, and complete multi-step tasks. OpenAI's enterprise strategy increasingly centers on agents that can perform work rather than simply generate responses, a focus the company has reinforced through agent execution environments and enterprise integrations . Google has expanded its own enterprise strategy through Gemini Enterprise and its enterprise agent platform , emphasizing integrations with workplace applications, enterprise data systems, and workflow tools. Large AI providers are increasingly building systems intended to handle software development, analysis, research, and workflow execution inside organizations. That context explains why Anthropic's launch partners skew heavily toward enterprise infrastructure rather than consumer products. Fable Enters the Enterprise AI Race Fable 5 launched across GitHub Copilot, Snowflake Cortex AI, cloud platforms, and enterprise AI environments. The Snowflake connection is notable because the company has spent years building enterprise AI infrastructure, including its own Arctic large language model and Cortex platform. These systems sit closer to enterprise operations than consumer AI products, giving organizations the ability to deploy models within software developme

How Anthropic Turned a Restricted AI Model Into an Enterprise Product

Keep reading. It's free.

Related stories