The White House is in advanced talks with OpenAI, Google, and Anthropic on a voluntary framework that would standardize how the most powerful AI models are reviewed and released, according to a Financial Times report published this week. An announcement could come as soon as next week, replacing the ad-hoc restrictions that have disrupted major model launches over the past two months.
Key Highlights
- The framework would set benchmarks for advanced models, define review timelines, and clarify which customers can access frontier systems inside the United States and abroad
- The Center for AI Standards and Innovation (CAISI) and the National Security Agency (NSA) are expected to play central roles in setting and monitoring the rules
- Negotiations have focused on two sticking points: the capability threshold that triggers a "frontier" review, and how long that review should take
- The deal would formalize and replace the improvised controls that delayed Anthropic's Claude Fable 5 and Mythos models and OpenAI's GPT-5.6
Details
The framework builds on Executive Order 14409, signed on June 2, which established the legal skeleton: classified capability benchmarks, a 30-day pre-release testing window, and voluntary participation by the leading labs. According to the reporting, an earlier version of the agreement was scheduled for signing in May but was canceled amid internal administration concerns that it was too restrictive; the review window was subsequently shortened.
The negotiations come after a turbulent two months for US model releases. The Commerce Department effectively banned Anthropic's Fable 5 and Mythos models over national security concerns in mid-June, then lifted the controls on July 1 — reportedly on condition that Anthropic deploy jailbreak filters blocking policy violations more than 99 percent of the time, the first known technical condition attached to a US AI release approval. OpenAI's GPT-5.6 launch was delayed at the government's request, with initial access limited to roughly 20 vetted partners approved personally by Commerce Secretary Howard Lutnick.
Google's Gemini 3.5 Pro, by contrast, launched in July without restrictions after scoring below the cybersecurity threshold that triggered reviews of its competitors' models.
Impact
For developers and enterprises, a formalized framework would bring predictability to a release process that has been opaque and inconsistent. The current case-by-case regime left customers unsure when — or whether — they could access newly announced models, and left labs negotiating individual customer approvals with the Commerce Department.
Critics warn the arrangement amounts to regulation without legislation. One Fortune analysis described the government as "repurposing existing legal authorities into what is effectively a backdoor licensing regime," while security researchers quoted by SecurityWeek said government access reviews "should not become the long-term default" for AI releases.
Background
The United States has never had binding federal rules for releasing AI models. The Biden-era voluntary commitments of 2023 covered testing and watermarking but carried no release-gating mechanism. The 2026 controls marked the first time Washington directly held back commercial AI products, and the resulting confusion — documented in reporting by The Hill — pushed both the administration and the labs toward a standing agreement.
What's Next
If the framework is announced on the reported timeline, the first models reviewed under it would likely be the next frontier releases from the three participating labs. Open questions remain: whether the classified benchmarks will ever be published, whether smaller labs and open-weight developers will be pulled into scope, and whether "voluntary" participation can survive a dispute between a lab and the NSA over a flagged capability.