on june 9, anthropic officially unveiled two new ai models—claude fable 5 and claude mythos 5—marking a pivotal step in its technological roadmap, shifting from general-purpose intelligence toward deep specialization for specific domains. fable 5 is open to the general public and has been officially designated as the most capable version currently available within the claude series; mythos 5, positioned as a “specialized model for highly sensitive fields,” is being delivered on an exclusive basis through the project glasswing initiative to cybersecurity defense agencies and critical infrastructure providers.
although both models are built on the same core architecture, anthropic has equipped them with distinctly different security policies and capability boundaries: fable 5 delivers significant advancements across software development, knowledge-intensive tasks, multimodal visual analysis, and research assistance, excelling particularly in long‑term, high‑complexity workflows and setting a new record for autonomous operating duration within the claude family. testing shows that it outperforms its predecessor, opus, in financial logic reasoning, deep understanding of unstructured documents, semantic reconstruction of charts and tables, and cross-domain problem decomposition. moreover, it can accurately extract quantitative parameters from specialized scientific images such as microscopic slides and spectrograms, and even reverse‑engineer fully functional front-end source code solely from web page screenshots.
to rigorously manage risks, fable 5 incorporates an upgraded content classification and interception system that proactively blocks high‑risk requests involving cyber defense, synthetic biology, hazardous chemicals, and model reverse engineering, automatically reverting to responses provided by claude opus 4.8. meanwhile, mythos 5 is tailored specifically for extremely demanding scenarios, demonstrating industry‑leading performance in cutting-edge areas like cybersecurity threat modeling, targeted drug molecule generation, protein structure prediction, and genomic function annotation. internal tests reveal that employing mythos 5 accelerates protein design workflows by nearly tenfold, with nine out of fourteen target proteins already yielding candidate structures exhibiting strong binding activity. in blind testing, researchers adopted up to 80% of the molecular hypotheses generated by mythos 5.
pricing is uniformly set at $10 per million input tokens and $50 per million output tokens. currently, fable 5 is fully accessible via api, while mythos 5 remains strictly limited to the glasswing partner ecosystem and selected bioinformatics research teams.
