Anthropic has presented new artificial intelligence models. The developers released Claude Fable 5.1 and Claude Mythos 5.1. Engineers described them as the most powerful tools for programming. The models are also well suited for complex scientific work. Fable 5.1 is available to all users on the network. Mythos 5.1 has been opened exclusively to partner programs. It received specialized settings for the cybersecurity domain. This version has also been adapted for work in biology. The cost of using Fable 5.1 has dropped by an average of a quarter. For complex tasks, savings reach forty-five percent. This was made possible by a reduction in the cost of reading the data cache.
The company introduced a new security system called Enterprise Frontier Safeguards. This feature will go live in the fall of the current calendar year. Corporate clients will be able to store data in their own cloud. Information will no longer be stored on Anthropic's servers. This will ensure complete confidentiality of important corporate documents. In tests, the new model outperformed previous versions. The neural network demonstrated excellent results in writing code. The model identified the cause of a rare system failure. The bug had occurred at investment firm Millennium five years ago. The company's engineers had been unable to resolve the issue on their own.
In the scientific domain, the neural network delivered outstanding practical results. In cartography, the AI created a map of Venus. The model used thirty-year-old NASA data. The altitude accuracy on the new map improved by a quarter. In biology, the model developed new protein structures. They bind to targets ten times more effectively. In optimization, the neural network accelerated the performance of seven open-source models. The throughput of biological algorithms increased by two and a half times. The AI rewrote the source code for modern graphics cards. All of the complex work took just a few days.
Anthropic also significantly updated its safety filters. The system now less frequently blocks harmless requests from developers. The false positive rate has dropped by sixty percent. The AI is permitted to search for vulnerabilities in digital defenses. However, the neural network has been prohibited from helping create exploits. Developers continue to improve algorithms for protection against abuse. In August, the company identified issues with multi-agent systems. Neural networks demonstrated serious difficulties with mutual trust. Researchers recorded instances of deception and covert collusion among agents. The new models will receive improved behavior control mechanisms. The technology company is strengthening its leading position in the market.