Washington didn't approve GPT-5.6. It just decided who got to use it first, tested it before anyone else could, and signed off before the company moved.
OpenAI's own account and the White House's account of the past two weeks tell almost the same story, right up until the word "approval." An executive order signed June 2 set up a voluntary system: frontier labs could hand the government a model for national-security review, for up to thirty days, before releasing it to trusted partners. Read literally, that window closes at the trusted-partner stage, not at public release. The order's own text is explicit that none of this creates a formal licensing or preclearance requirement. GPT-5.6 Sol, Terra, and Luna went into a restricted preview on June 26, which is the trusted-partner release the order's window is actually calibrated to. On the order's own terms, that is where the formal mechanism ends. The preview itself was limited to roughly twenty partner organizations OpenAI shared directly with federal officials, widely described in reporting as "government-vetted" or "government-approved" partners. Sam Altman told employees the government was approving that access customer by customer, and Commerce Secretary Howard Lutnick reportedly sought Altman's assurance that every relevant agency had signed off before OpenAI went wider. Axios, citing a source familiar with the review, reported that the broader release followed additional testing and meetings between OpenAI and federal officials. Other outlets repeated that account, but the specific testing arrangements have not been independently detailed by CAISI or OpenAI. According to Axios, CAISI participated in additional testing during the restricted-preview period.
Then, hours before the broad rollout, a White House official pushed back hard on how that sequence was being described. No green light was given, the official told reporters. No approval, no clearance. "No such permission is required or granted," the statement read, and release timing rests entirely with the company. GPT-5.6 went broadly available on July 9 regardless, across ChatGPT, Codex, and the API, thirteen days after the restricted preview began.
Both accounts can be true at once, and that is the actual story. The government did not issue a release license, because the order's own text rules that out. But it reviewed the model before its trusted partners got it, vetted the specific organizations on that list, and by OpenAI's own internal account, approved that access one customer at a time. Whatever shaped the thirteen days between that trusted-partner release and the public rollout was not the formal review window itself, since that window had already closed once the preview began. It looks more like sustained informal engagement, running on the same relationship the formal process created. Call that whatever you want. It is not what most people mean when they call a process voluntary, and it is not something the order's text actually accounts for either.
It helps to keep three separate government mechanisms apart here, because blending them is where the accuracy risk lives. There is voluntary pre-release review, which is what the June 2 order formally created, capped at thirty days and measured against release to trusted partners rather than the public — the mechanism both OpenAI and the White House agree covers the June 26 preview, and only the June 26 preview. There is export-control restriction, which is a different and more coercive tool, and which is exactly what hit Anthropic's Fable 5 and Mythos 5 in June under a Commerce Department order following a reported jailbreak-adjacent vulnerability. And there is informal government pressure over access and timing, which does not require either of the first two tools to still shape outcomes. GPT-5.6's story runs almost entirely on the third mechanism, dressed in the language of the first.
The pattern across both labs this year is not that government is claiming formal authority it does not have. It is that formal authority and practical influence have started to look identical from the outside, without anyone settling which one actually governs. Anthropic lost access outright under a restriction with real legal teeth. OpenAI kept nominal control the entire time and still ended up with the government choosing its early customer list. Different mechanisms, same result: neither company fully owned when or to whom its most capable model became available.
For enterprise security and procurement teams, the practical lesson isn't which lab handled this better. It's that flagship model availability is now a policy-sensitive dependency regardless of which access model a vendor has built. Procurement timelines, model-routing plans, and approved-model inventories need to account for the possibility of federal review activity delaying or reshaping access, even when no public rule, repeatable standard, or statutory approval process governs that activity. The government shaped this release at the distribution layer. OpenAI controlled identity and entitlement. Customers controlled what they did with access once they had it. No single party owned the full decision chain, and every party in it could still alter the outcome. One more thing worth knowing if your legal team goes looking: the order itself states it creates no enforceable right or benefit for any outside party. It cannot be used to compel a vendor's disclosure, and it gives no enterprise a legal claim if a vendor skips the voluntary process entirely. Whatever protection exists here runs through relationship and negotiation, not statute.
Closing Signal: That is the same governance gap this week's deeper Weaponized Access analysis, "The Common Gatekeeper," works through in full — what happens when neither a lab's access architecture nor a government's stated authority fully accounts for who actually controls the release.
