OpenAI to Release GPT-5.6 to 20 Partners After White House Ask

OpenAI will first release its next major model, GPT-5.6, to 20 trusted partners through Amazon’s Bedrock platform, after the Trump administration asked the company to stagger the rollout. The request, conveyed at a Wednesday staff meeting, marks the second time in two weeks that a frontier US AI lab has been told to gate its most powerful system. It shows how a pair of earlier government actions are quietly reshaping how cutting-edge models reach the public.

A person familiar with the meeting said Chief Executive Officer Sam Altman told employees the US government had asked OpenAI to keep GPT-5.6 inside a short list of trusted partners before opening it more widely. Altman stressed that staff need to work with the administration on any input officials may have on safety and restrictions, even where the company disagrees. OpenAI declined to comment. A White House official said the administration “continues to collaborate with frontier AI labs to develop shared approaches for addressing the challenges of scaling this technology.”

The Request and the 20 Partners

At the Wednesday all-hands, Altman told staff the government had asked for a phased GPT-5.6 release. Initially, the model will reach 20 partners, with access running through Amazon’s Bedrock software.

A person familiar with the meeting said Altman stressed that OpenAI staff need to work with the Trump administration on any input officials may have on safety and restrictions for upcoming models, even where the company disagrees. OpenAI declined to comment. The Information first reported the meeting. The 20-partner figure aligns with the kind of “trusted partner” tier the executive order describes.

The constraint lands on a model OpenAI has been racing to deploy. Under the new arrangement, the broad public release will now follow an initial phase with approved partners. That phase mirrors the pre-release review window the White House has been negotiating with frontier AI labs since the June 2 executive order. The order’s voluntary framework is being designed over the next 60 days. Until it lands, every release is effectively an ad-hoc negotiation between a lab and Washington.

• 20 trusted partners for GPT-5.6’s first wave
• 30 days maximum government pre-release access for covered frontier models
• 60 days window for Treasury, War, and Homeland Security to design the voluntary framework
• June 2, 2026 the executive order was signed

The Anthropic Precedent

The OpenAI request came in the shadow of an earlier directive that pulled Anthropic’s Fable 5 and Mythos 5 offline earlier this month. Anthropic said it received the directive at 5:21 p.m. ET, with no specifics on the national security concern. The action was framed as an export-control order requiring the company to suspend all access to the two models for any foreign national, whether inside or outside the United States.

Anthropic complied, but pushed back on the justification. The company said it had reviewed the government’s underlying report and concluded the level of capability displayed was widely available from other models, including OpenAI’s GPT-5.5. Anthropic called the jailbreak finding “narrow” and non-universal, and laid out the defense in a statement on its website. The episode set the practical precedent the OpenAI request now follows.

We are complying with the government’s legal directive and are removing access to Fable 5 and Mythos 5 for all users. However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.

Anthropic posted the statement the same day it received the directive.

The Executive Order Behind the Ask

The federal government’s authority over frontier-model releases rests on the full text of Executive Order 14409, signed by President Donald Trump on June 2, 2026. The order frames itself as pro-innovation, declaring that the United States “refuse[s] to stifle this innovation with overly burdensome regulation.” It also sets up the machinery now being applied to GPT-5.6.

Section 3 of the order directs the Treasury, War, and Homeland Security departments to design a voluntary framework under which AI developers can give the federal government access to covered frontier models for a period of up to 30 days before they plan to release such models to other trusted partners. The framework must also spell out how the government collaborates with developers to choose which trusted partners see the model first.

The order uses “covered frontier model” as a defined term, with the designation made by the NSA director in consultation with the cyber director and other agencies. The model qualifies once a classified benchmarking process shows its cyber capabilities cross a defined threshold. Until that benchmarking process is set up, the designation is happening case by case. That is the gray zone GPT-5.6 is now moving through.

The order gives federal agencies 60 days to complete the framework design. In the interim, the White House is convening frontier AI labs in meetings to work out the details of how models will be evaluated and which need review. The voluntary framework is meant to give agencies time to gauge what threats new models may pose to sensitive financial, national security, and critical systems. OpenAI’s staged GPT-5.6 release is functioning as a real-world test of the approach.

Who Counts as a “Trusted Partner”

For GPT-5.6, the trusted-partner tier runs through Amazon’s Bedrock hosting platform. The model will reach the 20 partners through that hosted channel rather than ChatGPT’s consumer surface. Bedrock’s enterprise customer base gives OpenAI a ready-made pool of vetted users. The arrangement also gives the White House a single, auditable choke point to track which partners get access and on what terms.

Bedrock as the gateway lines up with the structure of the executive order, which calls on developers to collaborate with the federal government to select trusted partners for early access. The 20-partner ceiling is not in the order itself; it appears to be a working figure from the OpenAI discussions.

Beyond the 20, the question is when GPT-5.6 opens more broadly. OpenAI has not given a timeline for lifting the trusted-partner gate. The arrangement looks designed to let the order’s 30-day pre-release window run first, with the broader public launch coming after. Customers outside the 20 will see the delay in practice. Anthropic’s experience offers a warning: Fable 5 has been offline since the directive, and access has not yet been restored.

OpenAI’s Position and the Altman Meeting

Altman met with members of the Trump administration at the White House in early June, days after the executive order was signed, according to an OpenAI spokesperson. He also sat down with Republican lawmakers during the same trip. The meetings came as the administration was laying out how it expected frontier labs to coordinate with the federal government before releasing new models. They were the first signal that OpenAI intended to comply with the new framework rather than push back. The collaboration continued in the weeks that followed. By Wednesday, the arrangement had produced a concrete decision about GPT-5.6.

At the staff meeting, Altman told employees the US government has grown more anxious about the capabilities of the most cutting-edge models. He stressed that staff need to work with the administration on any input officials may have on safety and restrictions, even where OpenAI disagrees.

Some OpenAI employees are concerned about how the government’s actions against Anthropic could affect the ChatGPT maker’s ability to roll out technology widely. Anthropic’s compliance with the Fable 5 and Mythos 5 directive cut those two models off from every customer, not just foreign nationals. OpenAI’s narrower compliance, limiting initial access to 20 partners, shows a lab choosing the least-disruptive path it can still defend. The decision frames future US frontier releases as negotiation outcomes, not unilateral launches.

Under the order, the 30-day pre-release access period is voluntary, not mandatory. The order’s Section 3(c) explicitly states that nothing in the section authorizes a mandatory licensing, preclearance, or permitting requirement. The collaboration is structured as a request the government makes and labs comply with. That is why the White House can press OpenAI on GPT-5.6 without issuing a formal directive.

Ball Saw It Coming

One person watching the shift with particular clarity is Dean Ball, a former Trump administration AI policy adviser who recently joined OpenAI. In mid-June, Ball posted on X that “AI is licensed now, but the requirements change constantly and are always a secret, even to the administration itself.” The post laid out a process the administration was running in real time, not a finished framework. Ball wrote it after Fable 5 went offline but before OpenAI faced its own confrontation with the government.

A week later, Ball updated his view. He posted on X that the whole US AI industry was “effectively frozen from new public releases until USG resolves the Fable situation they have stumbled into.” The prediction came before GPT-5.6 was visibly delayed. It also caught the larger shape of the moment: a frontier-model release pipeline in which the US government has the practical ability to slow, gate, or block a launch without issuing a new rule. Ball’s two posts together describe the ad-hoc licensing regime the OpenAI ask is now operating inside.

Ball’s move from the White House to OpenAI illustrates how thin the line between policymaker and lab has become. The same person who helped design the federal engagement framework is now helping OpenAI navigate it. That is one reason the two sides can move this fast on a 20-partner release. It also explains why OpenAI’s public posture has been collaboration, not litigation: the labs that helped draft the order are now the ones complying with it. OpenAI’s hiring of Dean Ball from the White House ran alongside its confidential S-1 filing.

Anthropic’s Push to Restore Access

While OpenAI stages its release, Anthropic is in active talks with the US government to restore access to Fable 5 and Mythos 5. The company has not been told when the suspension will lift. Anthropic argues that the underlying capability the government cited is widely available from other models, and that pulling a model deployed to hundreds of millions of users over a narrow jailbreak finding sets an impossible standard for the industry. The two sides remain apart. The outcome of those talks will set the floor for how broadly GPT-5.6 eventually opens.

The voluntary framework the order requires must be in place within 60 days of June 2. The administration’s frontier AI lab meetings are filling the gap in the meantime, and the OpenAI ask is the most concrete result so far. What happens with Fable 5 will decide whether the staged GPT-5.6 release is a one-off or the template.