The Week AI Became a National Security Asset: GPT-5.6, Government Vetoes, and the US-China AI Cold War

Something shifted last week. Not in the models themselves — though GPT-5.6 is genuinely impressive — but in the relationship between artificial intelligence and the state. For the first time in history, the world watched as two superpowers simultaneously treated frontier AI the way they treat nuclear technology: too dangerous to share, too powerful to release without oversight.

The sequence of events was extraordinary. On Thursday, OpenAI launched GPT-5.6 — its Sol, Terra, and Luna model family — after a multi-week delay imposed by the Trump administration over national security concerns. Days earlier, Anthropic had finally restored access to its Fable and Mythos models following a temporary export ban that blocked foreign nationals from using them. And on Monday, Reuters broke the story that Beijing had held meetings with Alibaba, ByteDance, and Z.ai about restricting overseas access to China’s most advanced AI models.

Three stories. One unmistakable signal: the AI Cold War has begun.

The Government Learns to Say No

The GPT-5.6 delay wasn’t a bureaucratic snafu. It was a deliberate test of a new regulatory muscle — the government’s willingness to stop a frontier AI release until it was satisfied the risks were manageable. The Department of Commerce, working with the Center for AI Safety, conducted additional testing on GPT-5.6’s cybersecurity capabilities before giving OpenAI the green light. The concern wasn’t theoretical. Advanced AI models can generate exploit code, automate social engineering attacks, and accelerate the discovery of software vulnerabilities. A model powerful enough to write production-grade software is also powerful enough to break it.

The same logic applied to Anthropic. The export ban on Fable and Mythos — however temporary — established a precedent: if your AI is good enough, the government can and will restrict who gets to use it. This isn’t export control for semiconductor fabrication equipment anymore. It’s export control for intelligence itself.

Beijing Mirrors Washington

While the U.S. was building its AI gatekeeping apparatus, China was doing the same — quietly, through meetings rather than public orders. China’s Ministry of Commerce has been in discussions with the country’s top AI firms about curbing foreign access to their most advanced models, including unreleased ones. The pattern is unmistakable: both nations now view cutting-edge AI as a strategic asset that confers military and economic advantage, and neither wants the other to have it.

The irony is hard to miss. For years, Chinese AI companies built international market share by giving their models away for free. Open-source releases from Alibaba’s Qwen team and DeepSeek were downloaded millions of times globally. Now Beijing is signaling that the era of AI free trade is over. The models that matter most will stay behind national firewalls.

What GPT-5.6 Actually Delivers

Amid the geopolitical drama, it’s easy to forget the models themselves are remarkable. GPT-5.6 Sol — the flagship — nearly matches Anthropic’s Claude Fable 5 on the AI Index benchmark while costing roughly one-third as much per token. On coding benchmarks, Sol leads. Sam Altman told CNBC the model is 54% more token-efficient on agentic coding tasks than its predecessor. The three-tier strategy — Sol for premium workloads, Terra for balanced performance, Luna for lightweight deployment — mirrors what the cloud computing industry learned a decade ago: one size doesn’t fit all.

OpenAI also launched ChatGPT Work, an autonomous agent that operates on users’ files and applications. This is where the rubber meets the road. AI isn’t just answering questions anymore. It’s taking actions — writing code, managing files, executing multi-step workflows. Every capability that makes these agents useful also makes them dangerous in the wrong hands. That’s precisely why governments are paying attention.

The New Normal

The events of this week aren’t a one-off. They’re the template for how frontier AI will be governed going forward. Every major model release from a U.S. lab will now face some form of government review — not as an exception, but as standard operating procedure. China will continue tightening control over its AI ecosystem, treating model weights as strategic exports. The middle ground — where AI was developed openly, shared freely, and governed lightly — is disappearing.

Whether this is good or bad depends on your vantage point. From a security perspective, screening models with cyber-offensive capabilities before broad release is defensible. From an innovation perspective, every week of delay is a week a competitor somewhere else isn’t waiting. From a global perspective, the Balkanization of AI into U.S.-aligned and China-aligned ecosystems risks creating two incompatible technological stacks — one for each side of a new digital iron curtain.

What’s clear is that the age of AI innocence is over. The technology has become too capable, too strategically significant, and too dangerous to be left to companies alone. Governments have arrived at the table — and they’re not leaving.