Now loading...
Anthropic has unveiled Claude Fable 5, the latest addition to its AI model lineup, which has received significant upgrades to ensure general safety. This model represents a leap in capabilities, outpacing previous iterations in numerous AI benchmarks, including software engineering, knowledge work, vision, and scientific research. Notably, Fable 5 excels at long and complex tasks, making it one of the most advanced AI models available for widespread use.
However, the release of such a powerful tool is not without its concerns. To mitigate the risks associated with its use, especially regarding cybersecurity, the model is equipped with safeguards. Fable 5 is designed to redirect requests on sensitive subjects to Claude Opus 4.8, its less capable predecessor, for a safer response. While these safeguards may occasionally misinterpret benign queries—triggering in under 5% of sessions—the company is keen on refining these measures with the goal of reducing false positives.
For a select group of cyberdefenders and infrastructure experts, a more advanced version named Claude Mythos 5 is being introduced with fewer restrictions. This version will first be accessible through Project Glasswing, a partnership with the U.S. government aimed at enhancing cybersecurity capabilities. Mythos 5 is touted as the most powerful model for cybersecurity globally, and access will eventually expand through a trusted program.
The functionalities of Fable 5 and Mythos 5 hold the potential for significant advancements in various fields. Early implementations under Project Glasswing have demonstrated their ability to assist cybersecurity professionals in securing vital software systems. In the life sciences domain, these models have contributed to innovative research, including developing novel hypotheses and accelerating therapeutic development.
Both models will be offered at a competitive pricing of $10 per million input tokens and $50 per million output tokens, which is less than half the cost of the previous Claude Mythos Preview. This release is a strategic move towards making sophisticated AI tools accessible to a broader audience efficiently and securely.
Users testing Fable 5 have reported marked improvements in various areas, such as its performance in software engineering tasks. For instance, initial testing revealed that Fable 5 could complete extensive coding projects within days, a task that would typically require months for a human team. Its ability to process and provide high-quality analytical results has also impressed early users, indicating a substantial step forward in AI-driven engineering.
Fable 5 has demonstrated its prowess beyond coding, achieving top scores in document analysis and complex reasoning tasks, surpassing previous models. Its capabilities extend into vision tasks, outperforming all predecessors by completing challenges that require intricate visual recognition without external aids. The model’s focus and memory features allow it to navigate long context tasks effectively, even outperforming the older Opus 4.8 model in complex game scenarios.
In drug design, Mythos 5 has significantly expedited aspects of protein engineering, reportedly performing tasks that would usually require skilled human involvement. Researchers have noted its ability to independently design binding sites and handle various facets of the drug design process. Moreover, experiments have shown that Mythos 5 was able to propose novel scientific hypotheses that have been validated by external studies.
To address safety concerns, Anthropic has implemented stringent safeguards across its new models to combat potential misuse, particularly in the realms of cybersecurity and advanced biological research. The sophisticated capabilities of Mythos-class models necessitate these precautions to prevent malicious actors from exploiting these tools. The newly developed classifiers are designed to detect and block unsafe requests related to sensitive topics.
Alongside the launch of these models, Anthropic is refining its data retention policies to enhance privacy and security. A new 30-day traffic retention policy for Mythos-class models aims to safeguard against complex attacks and further refine the system’s response capabilities.
Current users of Claude Mythos Preview will have the opportunity to upgrade to Mythos 5 without delay. Over time, the access to Mythos 5 is expected to broaden, fostering advancements in cybersecurity and medical research through a structured trusted access program. This comprehensive strategy underlines Anthropic’s commitment to responsibly expanding access to powerful AI technology.
Both Claude Fable 5 and Mythos 5 are available starting today, with the latter restricted to Glasswing partners and select biology researchers initially. The models are poised for substantial demand, with preliminary access structures in place for developers to explore their full capabilities.
