Anthropic has launched its latest offerings, Claude Fable 5.1 and Claude Mythos 5.1, which developers claim are the most powerful tools available for programming and scientific research.

Fable 5.1 is accessible to all users, while Mythos 5.1 is currently available only to participants in partner programs. The latter features enhanced security settings specifically designed for applications in cybersecurity and biology.

The cost of using Fable 5.1 has been reduced by an average of 25%, with savings of up to 45% for more complex tasks. This reduction is attributed to the decreased expenses associated with data reading from cache.

Additionally, the company introduced the Enterprise Frontier Safeguards (EFS) security system, which will go live in the fall. This system will allow corporate clients to store their data in their own cloud environments rather than on Anthropic's servers, ensuring complete confidentiality.

In testing, Fable 5.1 outperformed its predecessors in coding and logic capabilities. Notably, the model was able to diagnose the cause of a rare system failure at Millennium, an investment firm, a problem that had stumped engineers for five years.

In the realm of science, the AI achieved the following results:

  • Mapping: The AI generated a high-resolution map of Venus using data from NASA that is 30 years old, improving elevation accuracy by 25%;
  • Biology: The model designed protein structures that bind to targets ten times more effectively than existing counterparts;
  • Optimization: The AI accelerated the performance of seven open biological models by 2.5 times within days by rewriting the code for graphics processing units.

Furthermore, Anthropic has updated its security filters, now blocking harmless requests from programmers searching for vulnerabilities 60% less often. The AI is permitted to identify security weaknesses but is prohibited from aiding in the creation of vulnerabilities.

It is worth noting that in August, Anthropic identified issues of trust, deception, and collusion among multi-agent AI systems.