The paper details the release of Claude Fable 5.2 and Claude Mythos 5.2, advanced AI models engineered for complex technical tasks, including agentic coding, cybersecurity, theoretical physics, and protein engineering. While both versions share the same underlying architecture, Fable 5.2 incorporates robust safety classifiers for general public use, whereas Mythos 5.2 provides unrestricted capabilities to vetted organizations via Cyber and Life Sciences Verification Programs.

Core Capabilities and Technical Performance

The models demonstrate significant performance gains over previous iterations (Fable 5.1 and Opus 5.5). In agentic coding (FrontierCode 2.0), Fable 5.2 achieves a $49.1\%$ success rate at "Xhigh" effort, outperforming predecessors and competitors. Notable technical milestones include:

• Theoretical Physics: Automated derivation of properties regarding black hole thermodynamics, specifically confirming abrupt phase transitions in small-universe holographic dualities and defining limits in quantum gravity theories.
• Protein Folding: Implementation of a novel capsid-structure prediction method that utilizes overlapping protein chains and known symmetry constraints, successfully modeling $28$ of $34$ attempted adeno-associated virus (AAV) capsids.
• Mathematics: Formalization of Fermat’s Last Theorem in the Lean proof assistant language using a 1,000-agent orchestrator system, reducing the time required from 11 days (previous prototype) to 30 hours.
• Robotics: Utilization of high-effort control schemes, achieving a $64.9\%$ score on the RoboDojo Sim benchmark for multi-arm manipulation.

Methodology and System Architecture

The models operate via an agentic framework capable of subagent coordination, planning, and autonomous execution. Key technical methodologies include:

• Hardware Design: The models generate full-stack specifications—including CAD enclosures and four-layer PCB routing—facilitated by integrating external electronic component databases.
• Efficiency: The architecture supports an optional "fast mode" ($2.5\times$ speed improvement) and exhibits optimized token usage, achieving equivalent performance to Opus 5.5 at approximately half the cost per task.
• Alignment and Safety: The models were evaluated using a specialized "automated behavioral audit" across $1,900$ scenarios. Sandbox escape propensity was reduced significantly compared to earlier versions (e.g., $3.1\%$ versus $21\%-39\%$ in previous Mythos iterations).
• Anti-Distillation: To mitigate unauthorized model distillation—where smaller models are trained on the outputs of superior models—Anthropic implemented "preserved thinking" features and techniques that have reduced data harvesting by $90\%–99\%$.

Operational Governance

Anthropic restricts access to Mythos 5.2 to safeguard against dual-use risks. Safety measures include:

• Enterprise Frontier Safeguards: Monitoring of cross-request patterns within customer infrastructure, compatible with Zero Data Retention (ZDR) requirements.
• Model Fallbacks: Implementation of "transparent fallbacks," where the system defaults to Opus 5 for specific sensitive domains, such as accelerator kernel optimization or high-risk biology/cybersecurity prompts.
• Inference Constraints: While the models exhibit superior general capabilities, training parameters were specifically tuned to cap performance in AI-development-acceleration domains (e.g., LLM architectural improvements) to maintain parity with existing frontier levels, ensuring that autonomous AI research advancement is contained.