OpenAI has released Sol, a new frontier AI model, after a government review process that remains largely opaque to the public, researchers, and even other AI companies. Eighteen months into the Trump administration, there is still no published standard for what a frontier model must clear before shipping to the public. The Department of Commerce's Center for AI Standards and Innovation is currently leading an ad hoc review process, but the specifics of how Sol was evaluated remain undisclosed.

What Happened to OpenAI's Sol, and Why Is It Controversial?

OpenAI rolled out Sol for wide public access after what the company describes as a U.S. government review. However, nobody outside a small circle of officials and executives can describe how the model was actually cleared. Sam Altman told CNBC that OpenAI's conversations included Secretary of Commerce Howard Lutnick, Treasury Secretary Scott Bessent, and national cyber director Sean Cairncross, but neither the experts who tested Sol nor their methodology have been disclosed.

The model is considered at least on par with Anthropic's Fable, which was briefly pulled from wider access earlier this year over jailbreak concerns and friction between Anthropic's leadership and the administration. OpenAI previewed Sol for the government and select outside users before wider release, but the identity of those users and how they were selected has not been made public.

Why Does the Lack of Clear Standards Matter for AI Safety?

Researchers who track federal AI policy say they cannot audit what actually happened before Sol shipped. Mina Narayanan, a senior research analyst at Georgetown's Center for Security and Emerging Technology, noted that while Anthropic acknowledged conversations with the government around Fable and described defense-in-depth measures including a jailbreak classifier, the actual back-and-forth between the labs and federal officials is not public.

An executive order published last month sets an early-August deadline for six cabinet agencies to agree on a formal evaluation framework, but the specifics remain undefined. Sriram Krishnan, who served as senior advisor for AI in the White House until last month after joining from Andreessen Horowitz, told the Financial Times there will be no FDA-equivalent for AI.

"There will not be an FDA for AI," stated Sriram Krishnan, former White House senior advisor for AI.

Sriram Krishnan, Former White House Senior Advisor for AI

How Does Political Influence Shape AI Model Releases?

The political backdrop is difficult to separate from the review process. Altman has reportedly offered as much as 5% of OpenAI's equity to the administration's Trump Accounts program, and OpenAI president Greg Brockman is the largest publicly known donor to Trump's mid-term political operation. By contrast, Anthropic's Fable was briefly restricted after the government forbade its use by foreign nationals, partly over jailbreak-enabled hacking capabilities and partly due to friction between Anthropic's leadership and the administration.

Andy Konwinski, who co-founded Databricks and Perplexity and now runs the Laude Institute, said he has yet to meet anyone, including inside frontier labs, who can describe the release process end-to-end. He argues that the current setup concentrates decision-making in a small number of firms and a handful of officials, with almost no visibility for the broader research community.

What Are Experts Proposing as Solutions?

There is significant disagreement among policy experts about how to fix the current system. Some propose different regulatory approaches, while others worry about the concentration of power in the hands of a few companies and officials:

• Open Commons Model: Konwinski wants something closer to an open commons, modeled on the FDA, NIH, or the national labs, where government, industry, and independent researchers reach consensus on safety questions.
• Third-Party Auditing: Dean W. Ball, a former Trump policy advisor who now works for OpenAI, has proposed third-party auditing organizations licensed by the government to evaluate frontier labs' safety practices.
• Broader Research Participation: Alignment researchers, interpretability researchers, and data specialists are not meaningfully part of the release decision today, according to Konwinski, and should be included in any formal framework.

At last week's Open Frontier conference, Two Sigma founder David Siegel described what he called a very bad situation: a small number of firms controlling the technology, the government evaluating it in secretive laboratories, and neither the public nor the scientific community having meaningful access. For Sol, that description already fits the facts on the record.

What Are the Real-World Consequences of This Opacity?

The immediate market implication is asymmetric leverage. Labs with strong personal ties to the administration face a lighter, faster path to release; labs without those ties face export restrictions and pulled launches, as Anthropic did with Fable. That asymmetry will shape which U.S. models reach customers first, which foreign markets they can serve, and how aggressively competitors invest in Washington relationships versus safety infrastructure.

The economics push against transparency. Ball has noted that AI companies need to recoup training costs quickly, before competitors close the gap, which creates pressure to ship fast and share little. Konwinski points out that legal obligations and fiduciary duties are baked into how these companies operate, regardless of individual intent.

Until the early-August deadline produces an actual framework, and possibly well after, the release of the most consequential software in the industry will run on private phone calls and unpublished criteria. OpenAI declined to share details of the government's process, pointing instead to Sol's safety card, which cites external evaluations from the UK AI Safety Institute, SecureBio, and Irregular. In a late-June blog post, OpenAI said it does not believe this kind of government access process should become the long-term default.