Anthropic Unveils Claude Fable 5.1 and Mythos 5.1, Boosting AI for Science and Security

Anthropic has announced the release of its latest AI models, Claude Fable 5.1 and Claude Mythos 5.1, positioning them as the world's most advanced tools for coding and knowledge work. These new iterations promise not only enhanced performance but also a significant step forward in how AI can contribute to scientific progress, particularly in research-intensive fields. The company is emphasizing improvements in cost-effectiveness, data privacy, and the precision of safety mechanisms, addressing key customer feedback.

Details: Enhanced Capabilities and Cost Efficiencies

Claude Fable 5.1 and Claude Mythos 5.1 are fundamentally the same underlying model, differentiated by their safeguard levels. Fable 5.1 is broadly available, while Mythos 5.1 is accessible through specialized trusted access programs, tailored for sensitive work in cybersecurity and life sciences. A major highlight is the estimated 25% cost reduction for typical workloads billed by token compared to Fable 5, with savings potentially reaching up to 45% for highly agentic tasks. This is achieved by reducing pricing on cache reads, which are essential for processing and storing previously analyzed inputs.

Data privacy is addressed through the introduction of Enterprise Frontier Safeguards (EFS). This new system offers complete privacy, akin to a zero data retention policy, by ensuring data is stored within customer-controlled cloud infrastructure, not Anthropic's. EFS will be rolled out in phases starting this fall, with eligible enterprise customers able to use Fable 5.1 with zero data retention until its full availability. Safeguards have also been refined to reduce false positives, with cybersecurity applications seeing a 60% reduction in flagged benign content. Notably, Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits.

A New Performance Frontier

Performance benchmarks indicate a substantial leap forward for Fable 5.1. In scientific research, it achieved a 52.6% accuracy on the Terminal-Bench-Science 0.1 benchmark, a significant improvement over Fable 5's 24.7%. For agentic coding tasks, Fable 5.1 scored 55.8% on Terminal-Bench 4.0, compared to Fable 5's 42.0%. The model also demonstrated strong performance in knowledge work, achieving 1853 on GDPval-AA v2, surpassing Fable 5's 1723. These improvements are attributed to the model's ability to avoid shortcuts and address root causes of issues, as exemplified by its success in identifying a rare system crash for investment firm Millennium that had eluded human engineers and other AI models for years.

Scientific Research Capabilities

Anthropic has highlighted the potential of Claude Mythos 5.1 in scientific discovery. In molecular design, Mythos 5.1 designed protein binders with binding affinities ten times higher than top designs in industry competitions, achieving a nearly 50% hit rate across 12 targets, a significant improvement over the typical 10–15% hit rate. For computational analysis, Claude Fable 5.1 was used to train a neural network that generated a new, high-resolution elevation map of Venus, revealing details down to two to three kilometers with 25% greater height accuracy than previous maps. In computational biology, Mythos 5.1 optimized deep learning models, speeding up inference by up to 2.5 times and cutting estimated GPU costs by 30–60%, a task that would normally require weeks of engineering effort.

Safety, Security, and Alignment Advancements

Anthropic emphasizes that advancements in AI capabilities must be matched by progress in safety, security, and alignment. Claude Mythos 5.1 underwent extensive testing for chemical, biological, and cyber risks. While its capabilities are greater than its predecessor, evaluations indicate it still falls within acceptable risk tiers, leading to the deployment of similar safeguards as Mythos 5 for research biology. For cybersecurity, Mythos 5.1 demonstrated strong capabilities, and Fable 5.1's safeguards were stress-tested extensively, with no critical-severity jailbreaks found. Agentic safety evaluations showed Mythos 5.1 refusing malicious requests at a comparable rate to previous models and exhibiting improved robustness against prompt injections.

Alignment testing revealed that Claude Mythos 5.1 is better aligned across most metrics than Mythos 5, showing a reduced tendency to attempt resource access outside its environment or use motivated reasoning. The model also exhibits a lower rate of reward hacking. However, Anthropic acknowledges limitations in assessing very long-context work and multi-agent settings. To enhance enterprise privacy and security, EFS allows customers to manage their data on their own cloud infrastructure, with human review handled by the customer by default. Furthermore, safeguards for biology and cybersecurity have been made more precise, reducing false positives for benign queries while still protecting against threats. Anti-distillation mechanisms have also been strengthened to prevent the illicit extraction of model capabilities.

Trusted Access for Specialized Applications

Claude Mythos 5.1 will be available through two trusted access programs: the Cyber Verification Program (CVP) for defensive security work and the Life Sciences Verification Program (LSVP) for professional research and development. The CVP will soon include Mythos-class models, while the LSVP, developed in partnership with the US government, is expanding to the broader life sciences community. Additionally, Claude Security, Anthropic's product for scanning codebases for vulnerabilities, is now powered by Claude Mythos 5.1, further leveraging its advanced capabilities for defensive security applications.

Compliance and Future Outlook

Anthropic has also committed to compliance with the EU AI Act's Code of Practice on Transparency of AI-Generated Content, implementing an invisible watermark on outputs from models released after August 2, 2026. This watermark, detectable via a detection API, does not affect output quality or user privacy. The release of Fable 5.1 and Mythos 5.1 signifies Anthropic's continued dedication to pushing the boundaries of AI performance while prioritizing safety, privacy, and responsible deployment across critical sectors like science, cybersecurity, and general knowledge work.