Anthropic has unveiled its next-generation AI models: Claude Fable 5 and Claude Mythos 5. The developers highlight that the defining characteristic of these models is their capacity to autonomously tackle intricate, multi-step tasks over extended periods without continuous user input.
Claude Fable 5 is now accessible to a broad user base, while the more potent Mythos 5 version will be employed within the closed Project Glasswing program, focusing on cybersecurity and biological research challenges. Essentially, Claude Fable 5 is an iteration of the Claude Mythos model that previously played a role in identifying and replicating vulnerabilities in major software projects such as Linux, macOS, and Firefox, and also contributed to the creation of functional exploits during cybersecurity research.
Anthropic asserts that these new models can independently execute projects that would previously necessitate the involvement of entire teams of specialists.
As an illustration, the company cites its collaboration with Stripe. According to Anthropic, the model assisted in migrating a substantial codebase, roughly 50 million lines of Ruby. The company states that a comparable manual effort could have taken several months.
Professor Ethan Mollick from the Wharton School of Business provided another compelling example. He described how he provided Claude Fable 5 with a 19-page technical specification for developing an analytical tool to process unstructured survey data. He reported that the model worked on this project for approximately nine and a half hours, ultimately producing a complete software solution without any further user intervention.
The developers have also placed emphasis on the model’s multimodal capabilities. As a demonstration, Anthropic showcased Claude Fable 5 autonomously playing the game Pokémon FireRed, relying solely on visual input from the screen.
The company claims that previous generations of AI systems were unable to succeed in such a task, even when provided with additional tools and textual prompts.
Concurrently with the launch of the new models, Anthropic has enhanced its safety systems. When requests involving vulnerability discovery, exploit development, chemical synthesis, or attempts to replicate AI model architectures are detected, the system can automatically switch the user to a less advanced, previous-generation model (Opus 4.8).
The company assures that this safeguard will be deployed in a limited number of instances and will be accompanied by user notification. Anthropic believes such measures are crucial for the safe introduction of more powerful AI systems capable of extended autonomous operation and near-human-free task completion.
About the Author
