Anthropic just pushed the artificial intelligence sector into a highly regulated new era. As the company prepares for an initial public offering after hitting a $965 billion valuation—officially moving past OpenAI’s $852 billion mark—CEO Dario Amodei published a sweeping essay titled “Policy on the AI Exponential.” He argues the days of transparency-first AI development are over. Coinciding with this push for strict government intervention, Anthropic released its most capable technology yet: Claude Mythos 5 and Claude Fable 5.
Amodei wants the United States to implement binding, “FAA-style” safety regulations and mandatory third-party testing. He targeted these rules specifically at frontier models trained with over 10^25 FLOPs, according to a recent policy breakdown. The framework also applies to companies generating over $500 million in AI revenue or spending over $1 billion on AI research and development.
The essay asks governments for the legal authority to block the deployment of models presenting severe risks across cybersecurity, bioweapons, loss of AI control, and automated research. The framework proposes severe civil penalties tied to a violator’s global annual revenue for deployment holds violations. Amodei also warned of structural labor displacement, acknowledging that AI will act as a general substitute for labor and calling for immediate workforce transition plans.
Anthropic paired this warning with the highly restricted launch of Claude Mythos 5. This unsafeguarded frontier model is locked entirely to vetted cybersecurity organizations and government partners. The UK’s AI Security Institute already evaluated Mythos 5 and noted its autonomous cyberattack capabilities.
For the general public, Anthropic rolled out Claude Fable 5. This safeguarded version relies on real-time safety classifiers to block high-risk queries, automatically routing those prompts to the older Claude Opus 4.8 model. Fable 5 is live now, priced at $10 per million input tokens and $50 per million output tokens. Amazon Web Services immediately integrated Fable 5 into Amazon Bedrock, opening the model up to enterprise developers.
What the $10^25 FLOP Threshold Means for Enterprise AI Competition
Amodei’s specific targets—the $1 billion R&D spend and $500 million revenue marks—create a distinct regulatory ceiling for the tech industry. Startups building smaller, highly efficient models avoid the proposed government holds entirely. By demanding third-party testing only at the extreme frontier scale, Anthropic is formalizing a regulatory moat around the industry’s heaviest spenders right as it files its own confidential S-1 prospectus.
Because Fable 5 immediately landed on Amazon Bedrock, AWS developers now get access to Anthropic’s top-tier reasoning without managing the raw, unsafeguarded capabilities of Mythos 5. The safety classifiers routing high-risk prompts back to the older Opus 4.8 model allow Anthropic to scale its newest technology commercially without violating its own proposed cybersecurity boundaries.