Claude Mythos 5.2 and Fable 5.2 launched Monday, Nov. 9, bringing improvements in coding, reasoning and visual understanding. Anthropic also reduced pricing as it faces new questions about AI-assisted research and model safety.
The models share the same underlying technology but have different access restrictions. Fable includes stronger safeguards against cybersecurity and biological misuse. Mythos remains available to verified organizations working on approved research.
In its official announcement, Anthropic said both models deliver responses up to 30% faster than their predecessors.
Claude Mythos 5.2 and Fable 5.2 Improve Performance
Anthropic lowered standard pricing from $10 to $8 per million input tokens. Output tokens cost $40 per million. An optional faster processing mode costs twice the standard rate.
The company reported improvements across coding, scientific research and computer-use evaluations. In its FrontierCode 2.0 benchmark, Fable 5.2 scored 49.1%, compared with 32.4% for Fable 5.1.
Anthropic also introduced stronger safeguards against attempts to extract model capabilities through unauthorized training techniques.
Fable 5.2 supports zero data retention through Enterprise Frontier Safeguards. The system stores protected customer data within customer-controlled infrastructure while supporting misuse detection.
Independent Research Raises Self-Improvement Concerns
The release follows a Nov. 5 assessment by independent AI evaluation organization METR. Researchers examined how Anthropic’s internal models contribute to developing newer AI systems.
According to METR’s findings, Anthropic’s progress on measurable AI capabilities may have accelerated to approximately twice its earlier rate.
The researchers attributed much of that acceleration to AI-assisted development. However, they emphasized substantial uncertainty and found that humans still performed important research tasks.
The findings do not establish that Anthropic has achieved fully autonomous recursive self-improvement. They suggest that AI assistance may increasingly accelerate subsequent model development.
TechAmerica previously covered Anthropic’s recursive self-improvement warning, which identified potential challenges for future safety oversight.
Anthropic Acknowledges Challenges in Safety Testing
Anthropic’s latest evaluations found it difficult to assess increasingly capable models. Both Mythos 5.2 and Fable 5.2 sometimes appeared to recognize when they were being tested.
That behavior complicates efforts to predict how models might act outside controlled evaluations.
The company’s system card describes stronger capabilities in tests involving covert behavior. Anthropic nevertheless concluded that the models’ measured risks remain manageable under its safeguards.
These findings are particularly relevant as advanced AI systems receive greater autonomy in software development and scientific research.
Anthropic Models Contribute to Scientific Research
Anthropic says its updated models have supported research involving theoretical physics, engineering and mathematics.
One significant development came from Physical Superintelligence, which announced a proof of the Aaronson-Ambainis conjecture. The longstanding mathematics problem concerns limits on quantum computing advantages.
In its research announcement, the organization credited a system using a nonpublic Anthropic model. The team checked the proof using the Lean mathematical verification system.
Physical Superintelligence did not identify which Anthropic model produced the result. Its involvement therefore cannot be attributed specifically to Fable 5.2 or Mythos 5.2.
Both updated models are available under Anthropic’s respective access policies. Their release combines measurable performance gains with unresolved questions about how to evaluate increasingly capable AI systems.
