White House Nears Deal With OpenAI, Google, Anthropic on Early AI Model Review

The Trump administration is in advanced talks with OpenAI, Google, Anthropic, Microsoft and Amazon over voluntary rules for early government access to the newest artificial intelligence models before their public release. An announcement of the agreement is expected this week, foreshadowed by an earlier demand that OpenAI limit the rollout of its GPT-5.6 model to a narrow group of government-approved partners.

What the Agreement Covers

The mechanism would let companies developing the most advanced models voluntarily submit them to the U.S. government, which would have up to 30 days to assess whether a given system poses a national security risk before it reaches a wider circle of trusted partners or the general public. Technical teams at the relevant agencies would be tasked with evaluating models for risks tied to biology, chemistry and the ability to generate cyberattacks.

The document explicitly states that nothing in the agreement creates a mandatory licensing system, prior approval requirement, or permitting process for developing, publishing or distributing new AI models. The government's role is meant to be advisory, flagging risks rather than blocking launches, though critics note that even a voluntary mechanism carries informal political pressure on companies that depend on government contracts.

Fight Over the Threshold

A key sticking point in the negotiations is what capability threshold triggers the obligation to submit a model for review. AI companies are pushing for a high bar that would only capture genuinely groundbreaking systems, not routine updates to existing models. Government officials, drawing on assessments of how fast AI capabilities are advancing, want a lower threshold that would cover more launches than model developers are proposing.

GPT-5.6 as the First Test Case

A preview of the new mechanism came in late June, when the administration asked OpenAI to limit the early rollout of GPT-5.6 to a small group of government-approved partners. It was the first time the White House directly influenced the rollout schedule of a new model before it reached a wider audience, a signal that the U.S. government now treats the sequencing and pace of model releases as a matter of national security policy, not just market competition.

The GPT-5.6 case also shows how informal government pressure can operate even before a formal agreement is signed. OpenAI agreed to restrict access even though no signed document existed at the time requiring it to do so, which observers say demonstrates the real leverage the administration holds over companies that depend on federal contracts and their relationship with Washington.

The Anthropic Dispute Backdrop

The negotiations are unfolding against the backdrop of a recent dispute over exports of Anthropic's Fable 5 and Mythos 5 models, whose export ban the U.S. government lifted in early July after roughly three weeks of suspension, once the company agreed to strengthen safeguards and work with the government on model release protocols. The episode showed that the administration is willing to reach for hard export-control tools even before the voluntary framework takes effect.

For the market, this means the release order of the newest models from OpenAI, Google and Anthropic may increasingly depend not just on technical readiness but on how quickly agreements are reached with the administration in Washington. For companies outside this group of five, including European and Polish firms building products on top of U.S.-made models, this could mean delayed access to the latest versions if government review extends the public launch timeline.

Details of the agreement are expected to be announced in the coming days, and its final shape, especially the definition of the "covered frontier model" threshold, will determine how much of future model releases actually fall under the new mechanism.

Sources: TipRanks (tipranks.com), Reuters via Yahoo Finance (finance.yahoo.com), FAQ.com.tw (faq.com.tw)