Restoration of Claude Fable 5, Gemini's Video Dev Engine, DeepSeek Speeds Up Speculative Decoding

Claude Fable 5 was made available to global users on June 9, but just three days later, the US Department of Commerce issued an export control directive on national security grounds, requiring Anthropic to suspend access. The controversy stemmed from Amazon researchers using the model to obtain network attack methods. Anthropic argued that other models, including its own Claude Opus 4.8, could achieve similar results.

After negotiations from June 12 to 30, Anthropic agreed to add extra guardrails: some cybersecurity queries would be intercepted and routed to the less capable Opus 4.8. On July 1, Claude Fable 5 access was restored, but only on Anthropic-controlled platforms; third parties like AWS resumed later. Previously, the more advanced Claude Mythos 5 had been made available to government and critical infrastructure partners in April and was restored earlier on June 26.

This incident is one of the most notable conflicts between US AI companies and the government this year, involving Anthropic, Amazon, Google, Microsoft, and OpenAI. Commerce Secretary Howard Lutnick described the negotiations as making 'significant progress,' but after restoration, some users reported performance degradation, especially in biology, which Anthropic has not officially addressed.

The new guardrails on Claude Fable 5 target cybersecurity queries, but user feedback suggests the model also degraded in answering basic biology questions, implying the guardrails may be too broad or affect internal representations. Anthropic has not provided technical details, but such side effects are common in model fine-tuning.

Anthropic maintains that any powerful model can find the same vulnerabilities, so restricting a single model has limited effect; the government insists export controls are necessary. This fragmented regulation may prompt other companies to adjust release strategies or increase compliance costs.

If post-restoration performance issues are confirmed, they could erode user trust in top-tier models and accelerate exploration of open-source alternatives or local deployments. Currently, these reports are anecdotal and require more independent verification.

In his latest newsletter, Andrew Ng emphasizes that AI-assisted programming should follow the principle 'human tokens are precious, AI tokens are cheap.' He advises developers to skip lengthy upfront specifications and have coding agents build prototypes directly, then adjust requirements based on results. For example, he often spends 10 minutes writing a rough spec, lets the agent spend 20 minutes building, then refines based on observations.

This approach avoids waterfall processes: specifications are no longer a gate but a living document refined through iteration. Ng notes that early prototypes can be entirely discarded because the code is cheap, while the human feedback gained—'human tokens'—is the real value.

To address coding agents' forgetfulness—where forgetting human instructions is more costly than forgetting self-discovered information—Ng advocates recording key decisions in a SPEC.md file and having the agent update it with each iteration. This ensures human tokens are not wasted while leveraging cheap AI tokens for exploration.

The Claude Fable 5 incident shows that AI regulation is moving from principled declarations to concrete actions, but may bring unintended performance side effects. In the future, pre-release safety evaluation protocols (such as government-required standards) may become routine, and developers need to consider compliance in design.

At the same time, Ng’s iterative philosophy reminds us that AI’s real leverage lies in timely injection of human judgment, not pursuing one-time perfection. Even in highly regulated environments, rapid prototyping and flexible adjustments can accelerate innovation, provided safety guardrails do not cripple core model capabilities.

Uncertainty remains: Is the performance degradation widespread and persistent? Will export controls extend to other modalities like video generation? Can coding agents' forgetfulness be fully overcome? These questions will shape the next steps in AI development.