The New Claude Model Is Hiding Something (Fable 5)

Summary

The episode covers Anthropic's launch of Claude Fable 5, a highly capable AI model with strong safety restrictions, along with the more powerful but heavily guarded Mythos 5 variant. Demos highlight visual reasoning, coding, and long-horizon task execution. Benchmarks show Fable 5 outpacing GPT-5.5, raising competitive questions for OpenAI. Pricing notes and compute limitations are discussed, with a potential shift to usage-based credits after June 22.

• Anthropic released Fable 5 with restrictions on biology, cybersecurity, chemistry, and model distillation; an unrestricted Mythos 5 is only for vetted partners.
• Demos included 3D world building, playing Pokemon/Factorio, real-time feature building from customer calls, and network traffic visualization.
• On coding and real-world task benchmarks, Fable 5 substantially exceeds GPT-5.5 and rivals human expert workers.
• The model's capabilities spark debates on AI safety, access governance, and recursive self-improvement.
• Fable 5 is included in subscriptions until June 22, after which it may require usage credits due to high compute demands.
• The hosts see this as a pivotal moment where the most advanced AI is not universally accessible, marking a new era for the industry.