What exactly did OpenAI discuss with the U.S. government before the launch of its new model GPT-5.6?

OpenAI's latest flagship model, GPT-5.6, was recently released to the public, but the outside world has no idea what criteria the U.S. government used to deem it safe for release, and OpenAI itself refuses to disclose details.
(Previous Context: Trump administration partially lifts restrictions on Mythos 5; OpenAI GPT-5.6 Sol only available to White House-approved clients)
(Background: Paying to avoid trouble? OpenAI proposes donating 5% equity to U.S. sovereign wealth fund; Senator calls for "50% AI tax")

Table of Contents

Toggle

• Twelve Days of Black-Box Secrecy
• Who's Overseeing? Nobody Can Say
• Personal Connections or Institutional Process?

Twelve days — that's how long OpenAI's latest flagship model, GPT-5.6 (Sol), was "held back" by the U.S. government between completion of training and public release. The outside world still has no idea what criteria the U.S. government used to determine it was "safe to release." Even employees at frontier labs say they don't know what the process looks like.

Sam Altman said on CNBC that the process involved multiple rounds of discussions with officials including Commerce Secretary Lutnick, Treasury Secretary Bessent, and National Cyber Director Cairncross. But who conducted the tests and what standards were used? OpenAI refuses to disclose details. An AI model that affects the entire public is subject to a review process that remains a black box...

Twelve Days of Black-Box Secrecy

GPT-5.6 Sol is OpenAI's most powerful general-purpose model to date, alongside the family's Terra and Luna. It is widely regarded as on par with Anthropic's Fable. Fable had previously made the White House nervous due to its capabilities and was once banned from public release, partly because of concerns that it could be jailbroken to carry out hacking attacks.

GPT-5.6 now faces the same logic of review. The evaluation is currently led by the Commerce Department's CAISI (Center for AI Standards and Innovation). Earlier this year, after weeks of internal infighting, an executive order was released requiring six cabinet-level agencies to formalize their review processes by early August. In other words, the current oversight mechanism, at best, is "a patchwork job."

Sriram Krishnan, former White House senior AI advisor and former a16z partner, told the Financial Times: "There won't be an AI version of the FDA." This statement seems to give the industry a pass, but it also effectively acknowledges that the U.S. has no institutionalized review mechanism — everything relies on ad hoc coordination.

Who's Overseeing? Nobody Can Say

The only thing OpenAI is willing to put in the open is the safety card for Sol. Simply put, it's a report listing third-party test results. This time, it cites external evaluations from the UK AI Safety Institute (UK AISI), the biosecurity organization SecureBio, and the red-teaming company Irregular. But this is just one voluntarily disclosed aspect; the actual internal government review discussions remain hidden from the public.

Andy Konwinski, co-founder of Databricks and Perplexity, argued on X that this issue is fundamentally problematic — who has the right to decide whether a model can go live, and who oversees it? There is no clear answer. He advocates for an "open commons" model, modeled after the FDA, NIH, and national labs, where safety, alignment, and explainability researchers and data experts sit at the review table, rather than relying on a few phone calls.

Dean W. Ball, former policy advisor to Trump and now working at OpenAI, wrote in a newsletter that future oversight should be handled by government-approved third-party audit organizations.

Personal Connections or Institutional Process?

On the other hand, if this "light-touch regulation" is built on personal relationships with officials, the uncertainty and perverse incentives it creates are far more dangerous than a slow but transparent system.

Altman has previously been reported to be willing to donate up to 5% of OpenAI's equity to a fund for the "Trump account" run by the government. OpenAI President Greg Brockman is also a known major donor to Trump's midterm political operations. It is hard for the public to separate these political activities from the government's lenient attitude toward Sol.

David Siegel, founder of Two Sigma, asked attendees at the Open Frontier conference to imagine a very bad scenario: a few companies control key technology, the government evaluates whether the technology can be used in secret labs, and the public and scientific community have no access to the decision-making process. Today, this may no longer be imagination — it is already happening...