The newest thing in artificial intelligence is not a model. It is a permission slip.
OpenAI has built GPT-5.6 Sol, its most powerful model yet, with major advances in coding, biology and cybersecurity. In the old Silicon Valley story, the company would release it, brag about the benchmarks, watch developers swarm the API and let the market decide what happened next.
That is not what happened.
OpenAI is initially limiting access to Sol to selected trusted partners after previewing the model to the U.S. government. The company says broader access is coming, while warning that this kind of government access process should not become the long-term default because it keeps the best tools away from users, developers, enterprises, cyber defenders and global partners.¹
In other words, the public may get the new model later. First, Washington gets a look. Then approved users get a turn.
That might be defensible as a temporary response to a dangerous technology. It is harder to see it as isolated after what happened to Anthropic.
A few weeks earlier, the Trump administration used export-control authority to restrict access to Anthropic’s Fable and Mythos models. Mythos was built for cybersecurity work and, by Anthropic’s account, was unusually powerful. Fable was the safer public version.