OpenAI’s latest flagship model, GPT-5.6 (Sol), was “held” by the U.S. government for 12 days between completing training and its public release; Sam Altman said on CNBC that the process involved multiple rounds of talks with officials including Commerce Secretary Lutnick, Treasury Secretary Bessent, and National Cyber Director Cairncross, but OpenAI refused to disclose specific details.
Sam Altman: The review involved multiple rounds of talks with the commerce secretary, treasury secretary, and national cyber director
According to reports, Sam Altman said on CNBC that the U.S. government’s review process for GPT-5.6 involved multiple rounds of conversations with officials including Commerce Secretary Lutnick, Treasury Secretary Bessent, and National Cyber Director Cairncross; the CAISI (AI Standards and Innovation Center) under the Ministry of Commerce is currently the institution leading the evaluation.
However, OpenAI refused to disclose who conducted the specific testing, what standards were used, and what the basis is for the review of an AI model that the public uses—leaving outsiders with no way to verify.
CAISI leads the evaluations; an executive order requires six cabinet agencies to establish review procedures by early August
According to reports, an executive order that took weeks of internal infighting to finalize requires six cabinet agencies to set up official AI model review procedures by early August 2026. This means that the review approach GPT-5.6 went through is, at most, a “makeshift” temporary arrangement rather than a formal system.
Former White House AI senior adviser Sriram Krishnan told the Financial Times, “There won’t be an AI version of the FDA.” Andy Konwinski, co-founder of Databricks and Perplexity, argued on X that there’s no clear answer to who has the authority to decide whether a model can go live and who is supposed to oversee it; he called for an “open consensus” model modeled on the FDA, NIH, and national laboratories, so that safety and interpretability researchers and data experts can truly participate in review decision-making.
Sol safety card cites external evaluations; government internal review remains opaque
Based on what OpenAI proactively disclosed, GPT-5.6 Sol’s Safety Card cites external evaluations from the following three third-party institutions: the UK AI Safety Institute (UK AISI), the biosecurity organization SecureBio, and red-team testing firm Irregular.
However, the internal review discussions and the specific evaluation methods carried out by the government are entirely unknown to the public; Dean Ball, a former Trump policy adviser and currently with OpenAI, wrote in a newsletter that in the future oversight should be handled by third-party auditing organizations approved by the government to establish a more transparent accountability mechanism.
Altman reportedly donates 5% equity to a government fund; Greg Brockman is the biggest known Trump donor-in-chief
According to reports, the following political connections have raised doubts among observers about whether there’s any transfer of benefits in the GPT-5.6 review: Sam Altman has repeatedly been rumored to be willing to donate up to 5% of OpenAI equity to a government fund for “the Trump account.” OpenAI CEO Greg Brockman is reportedly the best-known biggest donor for Trump’s mid-term election political operations.
Two Sigma founder David Siegel proposed a hypothetical scenario at the Open Frontier conference: a small number of companies control key technologies, while the government evaluates them inside secret labs that the public and the scientific community can’t access—so outsiders inevitably draw parallels to reality.
Frequently asked questions
How long did the U.S. government review GPT-5.6 Sol before its official launch?
According to reports, GPT-5.6 Sol was “held” by the U.S. government for 12 days between completing training and its public release; Sam Altman said on CNBC that the review involved multiple rounds of talks with the commerce secretary, treasury secretary, and national cyber director, but OpenAI refused to disclose the specific review standards and methods.
Which U.S. government agency is currently responsible for reviewing AI models?
According to reports, the CAISI (AI Standards and Innovation Center) under the Ministry of Commerce currently leads the evaluations; however, an executive order requires six cabinet agencies to establish official review procedures by early August 2026, meaning the current mechanism is a temporary arrangement and institutionalized review has not been completed.
What third-party institutions has OpenAI disclosed for its safety evaluations of GPT-5.6?
According to OpenAI’s proactively published Sol Safety Card, it cites external evaluations from the UK AI Safety Institute (UK AISI), biosecurity organization SecureBio, and red-team testing firm Irregular; details of the government’s internal review discussions were not disclosed.
