Two things happened this month that look like industry inside-baseball but quietly touch anyone building on AI. OpenAI’s GPT-5.6 went public on July 9 — but only after roughly a month of federal review, released first to a handful of government-approved partners before the general public got access. And Anthropic, fresh off having its most powerful models briefly walled off by export controls, said it will now prerelease frontier models to federal authorities for testing and, alongside Amazon, Microsoft, and Google, proposed a shared framework for scoring security flaws. The pattern is the point: new AI models increasingly clear a government review you cannot see before you can use them.
What actually changed
A federal cybersecurity order this year asked AI companies to voluntarily submit their most powerful models for government testing before public release. That is why GPT-5.6 launched in two stages, and why last month’s export-control episode froze access to Claude’s top models for nearly three weeks. Frontier AI is being treated less like ordinary software and more like strategic infrastructure — closer to chips or telecom — with the delays and uncertainty that come with government attention software never used to get.
Why a small business should care
You are not submitting anything to anyone. But if any part of your plan depends on a specific model showing up on a specific date — a client demo built around a capability, a launch timed to a vendor’s roadmap, a workflow that only works on the newest model — you now have a dependency you do not control. Release dates can slip weeks for reasons that have nothing to do with the technology being ready. Google’s own flagship slipped its promised window this summer, for entirely different reasons. The lesson rhymes: roadmap promises are not delivery dates.
The practical read
Build so a delayed or restricted model is an inconvenience, not a crisis. Run your critical work on mainstream, widely available tiers rather than the bleeding-edge model that could get held up in review. Keep your prompts and workflows portable enough to swap models. And never promise a client something that only works if a not-yet-released model ships on time — promise what works today.
The honest caveat
This is not a reason to distrust AI or to freeze. Government review is, on balance, a reasonable response to genuinely powerful tools, and most of these delays are measured in weeks, not quarters. The point is not fear — it is to hold your dependencies loosely enough that a headline about a review or an export rule is something you read with mild interest, not something that takes your week down.
The takeaway
Model release dates now depend on a review process outside your vendor’s full control. Plan around what is shipping and in your hands today, keep a fallback model for anything critical, and treat every “coming soon” as a maybe. The businesses that stay calm through this stuff are the ones who never bet the month on a date they did not control.