Testing Mythos and Fable, Moving Beyond SWE-bench, Nvidia's Open Contender

Anthropic released Claude Fable 5, a guardrailed version of its Mythos model. The model uses a classifier to screen each prompt, routing questions about cybersecurity, biology, chemistry, or AI model engineering to the weaker model Claude Opus 4.8 or outright rejecting them. In Anthropic's own applications (e.g., Claude Code), flagged prompts are automatically routed to Claude Opus 4.8, but the switch is logged separately, requiring evaluators to manually distinguish. When accessed via API, flagged prompts are directly rejected, and evaluators can choose to fall back to Claude Opus 4.8 or mark the task as failed.

More notably, Anthropic initially covertly degraded model performance for users detected to be engaged in LLM research, without notifying users. After strong opposition, the company reversed this decision and committed to transparency, but Claude Fable 5 still refuses to use its latest capabilities to help AI researchers. Additionally, all users must accept a 30-day data retention policy, leading some evaluators to refuse submitting proprietary prompts due to data security concerns.

The U.S. government followed suit, using Commerce Department authority to regulate technologies that may pose national security threats, imposing export controls on Claude Mythos 5 and Claude Fable 5, requiring licenses for any foreign national (including Anthropic employees) to use them. This led Anthropic to globally disable Claude Fable 5. Andrew Ng noted that this action made many countries (including U.S. allies) realize that the U.S. can suddenly cut off their access to AI models, accelerating discussions on AI sovereignty and ensuring uninterrupted access.

OpenAI CEO Sam Altman commented that Anthropic's fear-marketing strategy—claiming to have built a 'bomb' and selling 'bomb shelters'—actually increased the likelihood of government export controls. Andrew Ng explicitly stated that he does not believe Anthropic built anything like a bomb, nor does he think export controls are appropriate.

Andrew Ng drew parallels to prior examples in semiconductors and rare earths: after the U.S. restricted China's access to semiconductors, China accelerated its own development; after China threatened U.S. rare earth supplies, the U.S. accelerated finding alternatives. Now, U.S. companies and government can quickly restrict other countries' access to frontier AI models, greatly incentivizing other countries to invest in open-source alternatives. Although training frontier models is not easy, Ng believes 'we have crossed the Rubicon.'

Microsoft CEO Satya Nadella wrote an article emphasizing the importance of building a healthy ecosystem on top of frontier AI technology. Andrew Ng agrees and hopes these events will ultimately become constructive steps. He calls for a freer, more open world where research is freely shared, and laws and social norms shape a level playing field.