Anthropic unveiled its latest artificial intelligence (AI) models, Claude Fable 5.1 and Mythos 5.1, on September 1, 2026. The new systems hit the market with notable improvements in accuracy and efficiency, as well as significant differences in security layers, with Fable aimed at the corporate market and Mythos for more restricted and specialized applications.
Performance and efficiency gains with new models
Anthropic’s Fable 5.1 and Mythos 5.1 models promise a considerable advance in the operational efficiency of several technologies. The company’s internal tests show that Fable 5.1 achieved 55.8% accuracy in the Terminal-Bench 4.0 benchmark, while Mythos 5.1 surpassed the 60% mark.
In the Terminal-Bench-Science 0.1 test, focused on scientific research by AI agents, Fable 5.1 achieved 52.6% accuracy. This result represents a significant increase when compared to the 24.7% of Fable 5 and the 29% of Opus 5, indicating a leap in search capacity.
The company also made adjustments to the operating cost structure. The processing value was kept at $10 per million incoming tokens and $50 per million outgoing tokens. These revisions generated substantial savings for users:
- Average savings of 25% on common workloads.
- Reduction of up to 45% in workloads with much more complex flows.
- 75% cut in the value of cached reads.
New safeguards strengthen enterprise security
One of the most important innovations in Fable 5.1 is the introduction of Enterprise Frontier Safeguards (EFS) technology. This security feature allows companies and developers to operate artificial intelligence models without retaining data in cloud infrastructures.
EFS functionality is crucial to meeting the growing demands for privacy and compliance in corporate environments. Additionally, Anthropic reported that Fable 5.1 engines have been optimized to reduce false positives in software audits by up to 60%, improving application reliability.
Mythos 5.1 specialized applications and their restrictions
Mythos 5.1 stands out for its high performance in extremely specific research and development scenarios. This model was applied to the design of protein ligands, obtaining an experimental success rate of around 50% in 12 biological targets evaluated in the laboratory.
Another notable use of Mythos was in optimizing GPU kernels for bioinformatics, resulting in a reduction in inference costs by up to 60%. Due to its highly sensitive and powerful nature, access to Mythos 5.1 is restricted. Only partners accredited by Anthropic’s Cyber Verification Program (CVP) and Life Sciences Verification Program (LSVP) will be able to use it.
Next steps and availability of artificial intelligence models
Claude Fable 5.1 is now available to the general developer and business market. On the other hand, Mythos 5.1 remains limited to participants in Anthropic’s trusted access programs, reinforcing the company’s strategy of balancing broad availability of tools with security for highly specialized and sensitive uses. Anthropic demonstrates its commitment to leading the advancement of AI while prioritizing accuracy and data protection.
