The Trump administration is in advanced negotiations with OpenAI, Google, and Anthropic to establish a voluntary pre-release review framework for frontier AI models, sources have told the Financial Times and other outlets. The proposed system would give the U.S. government up to 30 days to assess a new frontier model for national security implications before it is released to the public — stopping short of mandatory approval authority, but establishing the federal government’s first formal role in AI model launches. An announcement was expected as early as the week of July 7, 2026.
What Happened
Negotiations between White House officials and senior representatives from OpenAI, Google, and Anthropic have entered their final stages. The framework draws on a June 2026 executive order on AI and cybersecurity signed by President Trump that directed federal agencies to develop a review process for frontier AI systems. Under the proposed terms, participating AI companies would notify relevant government agencies before releasing a model that meets a certain capability threshold. Federal technical teams would then have a 30-day window to review the model for national security risks, including potential applications in biological, chemical, or cyberweapons development.
The government’s role during the review would be advisory — flagging concerns and recommending changes — rather than acting as a regulatory gatekeeper with the power to block a release outright. Companies that participate would gain access to a designated collaboration channel with trusted federal partners who could preview models under nondisclosure before public launch, an arrangement the White House is framing as a security benefit for participants as well as an obligation.
Why It Matters
The framework, if finalized, would represent a meaningful shift in the relationship between the U.S. government and the AI industry. For several years, frontier AI development has operated in an essentially ungoverned space when it comes to federal oversight — companies built and released models at their own discretion, subject to no mandatory pre-release review. The voluntary framework preserves that autonomy in legal terms while introducing a practical expectation that leading labs will consult with the government before major launches.
Critics argue that a voluntary framework has limited enforcement power: a company determined to move fast could simply decline to participate. Proponents counter that the reputational and regulatory cost of opting out of a widely adopted voluntary standard is itself a form of soft compliance — and that a voluntary framework can evolve into a mandatory one quickly if a safety incident occurs. Given the participation of OpenAI, Google, and Anthropic — who collectively represent the dominant share of frontier AI development in the United States — the framework could establish a de facto industry standard even without formal enforcement authority.
The 30-day review window has drawn pushback from the AI companies involved. In a market where being first to release a more capable model can shift developer adoption and enterprise contracts, a month-long pause represents a meaningful competitive cost. The negotiations have reportedly focused heavily on defining the capability threshold that would trigger the review, with AI companies pushing for the bar to be set high enough to exclude the majority of incremental model updates. Anthropic recently surpassed OpenAI on revenue, reaching $30 billion ARR — a milestone that underscores how commercially consequential model release timing has become for the participating companies.
Background and Context
The proposed framework arrives as AI governance conversations are accelerating on multiple fronts. Earlier this month, the United Nations convened its Global AI Governance Dialogue in Geneva, bringing together all 193 member states to discuss international approaches to managing frontier AI — a parallel multilateral process that reflects the degree to which AI policy has become a mainstream concern for heads of state, not just technology regulators.
At the national level, U.S. states have moved faster than federal policy in certain areas. California has been particularly active: the state recently secured a landmark agreement in which all California state agencies gained access to Claude AI at a 50 percent discount, integrating frontier AI into government operations while the federal framework is still being negotiated. That kind of state-level adoption creates additional urgency for a coherent federal governance posture, since government users are increasingly on the same frontier models that the administration is now seeking to review.
The Trump administration’s approach contrasts with the Biden era’s October 2023 executive order on AI, which required federal agencies to adopt safety standards but focused primarily on government procurement. The new framework attempts to position the federal government as a security partner to industry rather than a regulator, a framing consistent with the administration’s broader approach of reducing regulatory burden while maintaining national security prerogatives.
What Comes Next
If the announcement occurs in the week of July 7 as expected, the framework would likely require participating labs to sign letters of intent, after which the technical parameters — capability thresholds, review procedures, and agency roles — would be worked out over subsequent months. The precise threshold that triggers a review remains the central unresolved question: set it too low and every significant model update triggers a government review, slowing the release cadence of the world’s leading AI labs; set it too high and the framework provides only symbolic coverage of truly frontier-scale releases.
The ultimate question is whether the voluntary nature of the framework will hold as AI models grow more capable, or whether a significant safety incident will accelerate a push for mandatory pre-release oversight. What the current negotiations represent, regardless of their outcome, is a recognition by both industry and government that the ad hoc approach of the past several years is no longer tenable — and that some form of structured engagement between frontier AI developers and national security agencies is now a baseline expectation rather than an edge case.
