US Nears Voluntary AI Standards Deal: 30-Day Pre-Release Review, Classified Benchmarks, Meta Not at the Table
The Financial Times, relayed by Reuters on July 2, reported that the Trump administration is in advanced talks with major AI companies to create voluntary standards for frontier model releases, with an announcement possible as early as next week. The mechanics have been partially visible for weeks in how individual model launches have played out. The deal would formalise them.
What the Framework Covers
The reported agreement has four pillars.
Benchmark definitions. A classified process led by the NSA and the Center for AI Standards and Innovation (CAISI, under the Commerce Department) would determine which models qualify as “covered frontier models” — those with advanced enough cyber capabilities to require review. The threshold itself stays classified.
Timelines. Developers of covered models would submit them for a government review window before public release. The current executive order, signed by Trump in June, set that window at 30 days. Earlier proposals ran to 90 days; the shorter window was a concession to keep voluntary participation viable.
Access rules. The framework would clarify who can access covered models, both domestically and internationally. This addresses the ad-hoc export controls that took Anthropic’s Fable 5 and Mythos 5 offline for 19 days in June and constrained GPT-5.6 Sol’s launch.
No mandatory licensing. The June executive order was explicit: “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.” The current talks preserve that limit. Developers sign on voluntarily; there is no preclearance before release.
Evidence the Regime Is Already Running
The framework formalises enforcement actions that have already happened. Commerce suspended Fable 5 and Mythos 5 on June 12 after Amazon researchers surfaced a jailbreak producing functional exploit code; both models returned June 30 to July 1 following government review. OpenAI delayed GPT-5.6 Sol’s general availability at the White House’s request, restricting access to roughly 20 vetted organisations while cybersecurity reviews complete.
Google is separately in talks with the government ahead of an unspecified “advanced coding model with more sophisticated capabilities,” according to a Reuters source. The breadth of that model’s reach — and presumably its cyber risk profile — appears to be what triggered the conversation.
Meta Is Not Signing
The companies identified in current talks are OpenAI, Anthropic, Google, Amazon, and Microsoft. Meta is not included. Reporting from the week prior indicated the administration was “working overtime” to secure Meta’s participation, without success as of publication.
Meta is the only major US lab without a pre-release AI safety agreement of any kind. The company has also publicly opposed mandatory safety commitments. Its absence from a voluntary deal that every other frontier lab is joining will invite scrutiny as both OpenAI and Anthropic prepare for IPOs and want to demonstrate institutional alignment with government.
What Changes If the Deal Closes
A formal voluntary framework would give developers a defined process rather than case-by-case negotiations. Both OpenAI and Anthropic are currently navigating government constraints on an improvised basis — no published rules, no defined timeline, no clear appeal path. The deal would establish all three.
The cost of joining is a 30-day delay on any model that crosses the classified capability threshold. The cost of not joining is the current situation: opaque restrictions that have pulled a flagship model offline mid-quarter and delayed a competitor’s launch by an indefinite number of weeks.
The announcement window the FT cited — “as early as next week” — makes the week of July 7 the earliest plausible date for the framework to become public.