The United States is close to formalizing how the most capable AI models reach the market. The government is in advanced talks with OpenAI, Anthropic and Google on a voluntary framework of standards for releasing frontier models, the Financial Times reported in early July, with an announcement possible within days.

What's being negotiated

According to the reporting, the framework would set benchmarks for what counts as a "covered frontier model," define release timelines, and clarify who can access such models inside the United States and abroad. It operationalizes the June 2 executive order on advanced AI, which established a classified capability benchmark — built with input from Treasury, NSA, CISA and NIST — and a government review window of up to 30 days before public release.

Already operating in practice

The machinery has been visible for weeks. On June 12 the Commerce Department ordered Anthropic to cut off foreign access to its Fable 5 and Mythos 5 models — an 18-day shutdown lifted at the end of June. OpenAI delayed the full public launch of GPT-5.6 "at the U.S. government's request," per Reuters, limiting it to roughly 20 vetted partner organizations approved customer by customer; Commerce Secretary Howard Lutnick reportedly called Sam Altman personally to warn against releasing without government sign-off.

The criticism

Dean Ball, a former White House AI adviser, has been blunt: the arrangement amounts to "a de facto involuntary licensing regime," he argues — one that never went through Congress. "Nobody knows what the requirements are to get licensed, and the administration itself does not seem to know," Ball wrote. The criteria defining a covered frontier model remain classified, meaning models are being gated before the standard justifying the gate has been written down. Critics add that open-ended review clocks could delay US releases while Chinese developers operate outside the framework entirely.

What to watch

Open questions for the final text: whether Meta signs on, how "frontier" thresholds are defined, how the 30-day clock actually runs, and what access foreign allies get. A White House announcement, if the FT's timing holds, would land the same week the UN wraps its first global AI-governance dialogue in Geneva.