OpenAI, Anthropic push 30-day review for frontier AI models

OpenAI and Anthropic are working together in Washington to shape a federal review system for advanced AI models, despite competing for customers, funding, and technological leadership.

Summary

• OpenAI and Anthropic support up to 30 days of federal access before some frontier model releases.
• The proposed process would apply across the industry, including rivals such as Meta and xAI.
• Federal agencies must develop the framework by Aug. 1 under a June executive order.
• Nvidia, Meta, and Microsoft separately warned against sweeping controls on open-weight AI models.

OpenAI and Anthropic seek common review standards

OpenAI and Anthropic are urging the Trump administration to adopt a consistent review process for AI models with advanced cybersecurity or national security capabilities, according to The Information.

Their cooperation comes before an Aug. 1 deadline for federal agencies to define which systems should qualify as "covered frontier models." Those models could be provided to the government for evaluation for up to 30 days before release to other trusted partners.

Both companies reportedly want the standards applied across the AI industry rather than limited to developers already cooperating with Washington. That could bring competing companies, including Meta and Elon Musk's xAI, under the same review framework if their models cross the eventual capability threshold.

The discussions mark a rare policy alignment between OpenAI and Anthropic. Both companies sell general-purpose AI systems and compete for enterprise contracts, researchers, computing capacity, and investment.

Neither company has publicly released the full terms it wants included in the final framework.

Trump order sets voluntary 30-day process

President Donald Trump ordered federal agencies to establish a classified benchmarking process through Executive Order 14409 on June 2.

The order directs the National Security Agency, Cybersecurity and Infrastructure Security Agency, Treasury Department, and other federal bodies to determine when an AI model has capabilities strong enough to warrant additional review.

Developers participating in the framework would be able to ask the government whether a model meets the covered frontier model threshold. They could then provide federal evaluators with access for up to 30 days before making the system available to other approved partners.

The order requires safeguards for intellectual property, confidential information, cybersecurity, and insider threats during the evaluation.

However, it expressly states that the process cannot create mandatory federal licensing, preclearance, or permitting rules for developing or releasing AI models. Companies would therefore participate voluntarily unless separate laws or government powers apply.

US AI developers seek clarity before launches

A shared framework could give US developers clearer standards for deciding when government testing is expected. Questions remain over which technical benchmarks will trigger a review, which agencies will test each system, and how officials will approve early-access partners.

Those uncertainties have already affected recent launches. OpenAI and Anthropic provided government officials with early access to advanced models before wider distribution, but outside researchers had limited visibility into the evaluations.

A July report found that the agencies, tests, and approval requirements involved in those reviews remained unclear. The lack of a published process could make release planning harder for developers and leave smaller AI companies unsure whether the same standards apply to them.

Applying one capability-based threshold to OpenAI, Anthropic, Meta, xAI, and other developers could reduce some of that uncertainty. It could also limit concerns that Washington is giving selected companies different treatment.

Open-weight AI debate divides the industry

OpenAI and Anthropic's call for a common review process comes as other technology companies are warning Washington against placing broad controls on open-weight AI models.

As crypto.news reported last week, Nvidia, Meta, and Microsoft joined 22 other organizations in an open letter arguing that sweeping restrictions could weaken US leadership as competition with China intensifies. IBM, Palantir, Mistral, Hugging Face, Mozilla, Andreessen Horowitz, and the Linux Foundation also signed the letter.

The group called for targeted legal and commercial measures against misuse instead of restrictions covering technologies with legitimate research and business applications.

Open-weight models allow companies, researchers, and governments to download software, modify it, and run it on their own infrastructure. Supporters say this gives users greater control over their data, security, and computing systems while lowering deployment costs.

The signatories did not frame open and closed models as mutually exclusive. Instead, they argued that both approaches are needed to support competition and give developers different ways to build and deploy AI systems.

Nvidia CEO Jensen Huang shared the letter in his first post on X and defended the role of both development models.

"The world needs both frontier closed models and frontier open models."

Elon Musk also supported the letter in a reply to Huang, although his AI company, xAI, was not listed among the 25 reported signatories.

The two efforts address different parts of the policy debate. OpenAI and Anthropic are seeking consistent federal evaluations for models that cross a national security capability threshold, while the open letter argues against restrictions based mainly on whether a model makes its weights available.

Federal agencies now have until Aug. 1 to define the frontier-model threshold and outline the voluntary review process. The final framework will show whether Washington focuses on measurable capabilities or applies broader conditions based on how AI models are distributed.