Anthropic has launched its new model, Claude Fable 5, which is a rebranded and safer version of the previously discussed Mythos 5. While the full Mythos 5 model remains limited to select users under the Project Glass Wing program, Claude Fable 5 is generally available and claimed to surpass the capabilities of any prior publicly accessible Anthropic model. The release aims to allow broader testing and real-world application, particularly in coding and other complex tasks, though the presenter notes that thorough evaluation requires extended use to understand nuances like token efficiency and cost-effectiveness.
Pricing for Claude Fable 5 and Mythos 5 is higher than previous models, roughly double that of Opus 4.8, but notably less than initially feared and significantly cheaper than the earlier Mythos preview. This reduction is attributed to improved inference compute capacity, possibly linked to a deal with Elon Musk. Benchmark results show Claude Fable 5 outperforming Opus 4.8 and GPT 5.5 in many areas, including new benchmarks like Frontier Code, though gains are not uniform across all tasks. Some benchmarks, such as tool use and computer use, show only modest improvements, raising questions about whether the price premium is justified for all use cases.
A major focus of the release is on safety and content moderation. Claude Fable 5 incorporates strict safeguards that automatically flag and block requests related to cybersecurity, malware, life sciences, and attempts to extract long chains of thought from the model. These safety measures sometimes trigger even on relatively benign queries, as demonstrated by the presenter’s attempt to discuss Ebola outbreaks, which was redirected to a less capable model. Anthropic has implemented classifiers that monitor all inputs, including uploaded files, to prevent jailbreaking and misuse, reflecting a cautious approach to balancing capability with responsible deployment.
Another significant change is Anthropic’s updated data retention policy, requiring 30-day retention of all traffic on Mythos-class models, including third-party platforms. This marks a shift from previous practices where data privacy was a selling point, especially on cloud platforms like Google Cloud. Although Anthropic claims data will not be used for training new models or non-safety purposes, the presenter expresses skepticism about how data derivatives might still be leveraged. This policy change could impact customer trust and adoption, particularly among privacy-conscious users and enterprises.
Finally, the presenter highlights the evolving subscription and pricing model for Claude Fable 5. While early access is included in certain pro and max plans until June 22, after that date users will need to pay per token usage, signaling a move away from unlimited token plans. This shift may influence how individuals and companies choose to integrate Claude Fable 5 into their workflows, potentially using it for planning and higher-level tasks while relying on cheaper models like Opus or Sonnet for execution. The presenter plans further testing and invites feedback from users to better understand the model’s practical value and limitations.
