Blog

Approved Customer by Customer: What the GPT-5.6 Gate Means for AI Governance

Editorial graphic showing a lone figure between black gates labeled GPT-5.6 and Customer by Customer, with the U.S. Capitol behind one gate and a red access warning.

TL;DR

  • OpenAI’s next model, GPT-5.6, ships first only to partners the federal government approves, vetted one customer at a time.
  • The request follows Trump’s June 2 executive order calling for “voluntary” pre-release review of the most capable models, up to 30 days.
  • It also follows the June 12 export-control directive that forced Anthropic to pull Fable 5 and Mythos 5 worldwide, three days after launch.
  • No criteria or framework has been published. Agencies are clearing access case by case.
  • For builders, access to a hosted frontier model is conditional and can be cut off on same-day notice, with no appeal.
  • A process labeled voluntary is now working like a license.

This week OpenAI told its staff that the next model, GPT-5.6, will not reach the public the way its other models did. A small group of partners gets it first, and the federal government decides who makes that list, one customer at a time.[1] That is new. Sam Altman said it plainly in an internal memo: the government would be “approving access customer by customer during this preview period,” with a wider release perhaps “a couple of weeks later.”[1][2] The White House asked for the limited rollout because of the model’s advanced capabilities, a person familiar with the matter told CNN.[3] OpenAI has not posted an official announcement, and it has not said who qualifies.[2]

Three weeks earlier, on June 2, President Trump signed an executive order asking AI companies to hand their most capable models to the government for testing, up to 30 days, before any public release.[4] The same order tells federal agencies to build benchmarks for AI cyber capabilities and to stand up an “AI cybersecurity clearinghouse,” and it trimmed an earlier 90-day review window down to 30.[4] The word attached to all of it was voluntary. GPT-5.6 is the first real test of what that word means once a model is finished, and so far the answer is that a federal office clears the customer list before anyone outside the building gets in. Two offices drove the request: the Office of the National Cyber Director and the Office of Science and Technology Policy.[1] After OpenAI had already briefed senior officials, Commerce Secretary Howard Lutnick called Altman and warned the company not to proceed without sign-off from more agencies.[1] Altman is complying. He also wrote that this “is not our preferred long term model” and that OpenAI wants “a more sustainable approach for future releases.”[1][2]

None of this started with OpenAI. On June 12, at 5:21 in the evening, Anthropic received an export control directive ordering it to cut off access to its two most capable models, Fable 5 and Mythos 5, for any foreign national anywhere, including its own foreign employees.[5] The models had been public for three days.[6] You cannot reliably separate foreign nationals from US users in real time across a base that large, so Anthropic switched both models off for everyone.[5] The stated reason was a jailbreak. Anthropic reviewed the demonstration behind the order and said it surfaced a handful of already-known, minor flaws that other public models, OpenAI’s GPT-5.5 among them, could find just as easily.[5] The company complied and objected in the same breath, arguing that a narrow potential jailbreak should not pull a commercial model away from hundreds of millions of people, and that if the government is going to block a deployment it should do so through a process that is “transparent, fair, clear, and grounded in technical facts.”[5]

One wrinkle is worth keeping honest. Anthropic’s capability claims had been contested for months before the takedown. Independent researchers reported that cheaper, open-weight models could reproduce much of what Mythos did, and that the count of serious exploits was smaller than the marketing suggested.[7] The government never made its specific evidence public.[5] So a model came down over a risk that several outside parties believed was overstated, on the strength of a letter that named no standard and showed no proof.

Set the two episodes side by side and the shape is hard to miss. A review that is voluntary on paper is working as a gate in practice. No framework for it has been published; the government is still “expected” to build one.[2] The cyber benchmarks the executive order calls for do not exist yet.[4] OpenAI has not said which customers qualify or what the bar is.[2] So the decision about who can use the most capable American models is being made one case at a time, by agencies, with no written standard a company can read or contest. Call it what you like. It behaves like a license.

For anyone building on these systems, the Fable case taught the lesson in one evening. Access is conditional. The condition can change on same-day notice, with no appeal.[5] Teams that treated a rented model as fixed infrastructure learned otherwise on the twelfth.[8] Some are now looking harder at open-weight models they can run themselves, for the plain reason that a model on your own servers cannot be switched off from Washington.[8]

The real question here is about process. Almost no one argues a government should be powerless to stop a genuinely dangerous model, and Anthropic itself said the government should hold that authority.[5] What matters is the path that authority runs through: a clear statutory process with published criteria and disclosed evidence, or phone calls and letters that name no standard. Right now, it is the second. The written framework is supposed to come later.[4] Until it does, voluntary will keep meaning something close to approved, one customer at a time.

Sources

[1] Matthias Bastian, “OpenAI’s GPT-5.6 rollout now requires US government approval on a ‘customer by customer basis,’” The Decoder, June 26, 2026. https://the-decoder.com/openais-gpt-5-6-rollout-now-requires-us-government-approval-on-a-customer-by-customer-basis/

[2] “OpenAI will initially only release ChatGPT 5.6 to government-approved customers,” Engadget, June 25, 2026. https://www.engadget.com/2202129/openai-will-initially-only-release-chatgpt-5-6-to-government-approved-customers/

[3] “White House asks OpenAI to limit its next model release,” CNN Business, June 25, 2026. https://www.cnn.com/2026/06/25/tech/openai-limit-release-white-house

[4] “Trump signs AI safety order seeking voluntary review of new models,” NPR, June 2, 2026. https://www.npr.org/2026/06/02/nx-s1-5844347/ai-safety-trump-executive-order

[5] “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” Anthropic, June 12, 2026. https://www.anthropic.com/news/fable-mythos-access

[6] “Anthropic Disables Claude Fable 5 and Mythos 5 After US Government Order,” MarkTechPost, June 13, 2026. https://www.marktechpost.com/2026/06/13/anthropic-disables-claude-fable-5-and-mythos-5-after-us-government-order/

[7] “U.S. gov’t orders Anthropic to disable its newest AI models worldwide due to security threats,” Tom’s Hardware, June 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/us-export-control-order-forces-anthropic-to-disable-claude-fable-5-and-mythos-5-worldwide

[8] “When a Government Pulls an AI Model: What the Fable 5 and Mythos 5 Suspension Means for Security Teams,” Snyk, June 2026. https://snyk.io/blog/fable-mythos-suspension-security-takeaways/