Anthropic has updated its lineup of neural networks, unveiling the Claude Fable 5.2 and Claude Mythos 5.2 models. They are focused on solving the most complex technical tasks and during testing were able to independently design electronic devices and write an operating system from scratch.
As noted in the developers' official blog, technically these are one and the same model. The difference lies in the fact that Fable has strict safeguards in the areas of cybersecurity and biology, and is therefore available to all users on cloud platforms. Access to Mythos, by contrast, is granted exclusively to research organizations that undergo a special vetting process.
Particular emphasis in the new models is placed on programming. The neural network already generates over 90% of the changes in Anthropic's own codebase. During testing, a team of agents based on Fable 5.2 successfully migrated a mission-critical company service from Python to Rust, accelerating its startup by 26 times. In addition, at the request of one of the engineers, the artificial intelligence created a fully fledged operating system in eight days: the algorithm coordinated over 1,000 subagents, writing the kernel, drivers, file storage system, and graphical interface.
In addition to writing software, the models demonstrated the ability to design hardware. From a single text prompt, the artificial intelligence developed three working electronic devices: a home voice hub, a USB speakerphone, and a virtual pet. The neural network independently prepared specifications, selected components, designed four-layer boards and 3D enclosure models, and each board was ready for production in less than a day.
"To guarantee that safety keeps pace with the models' growing capabilities, we conducted rigorous testing that included developing new methods for assessing specific threats, such as sandbox escape. We believe that misalignment risks remain low, and Fable 5.2's safeguards create fewer unnecessary blocks," Anthropic emphasizes.
The new versions generate responses 30% faster than their predecessors. The base cost is $8 per million input tokens and $40 per million output tokens. The developers also added an optional "fast mode" that speeds up generation by 2.5 times, but it will cost twice as much.
