OpenAI Limits GPT-5.6 to Trump-Approved Partners in New AI Gatekeeping Era

OpenAI on Friday released GPT-5.6 Sol, its most powerful model yet, but only to a “small group of trusted partners” vetted through a Trump AI cybersecurity review. It is the second time in two weeks the new federal review process has reshaped a frontier launch.

The order is “voluntary” on paper. On June 13, the US government told Anthropic to pull Fable 5 and Mythos 5 offline, citing national security concerns over a potential narrow jailbreak. Anthropic complied but called the move a “misunderstanding” and warned it could halt all new model deployments.

Now OpenAI is following the same playbook, and signaling, in unusually blunt language, that the arrangement cannot hold.

The Executive Order That Set the Stage

President Trump signed Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” on June 2, 2026, putting the federal government on the front end of the AI release cycle. The order directs the Treasury Department, the Department of War, and the Cybersecurity and Infrastructure Security Agency, in consultation with the National Security Agency, to “develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models.” When a model crosses a classified threshold, it gets the designation of “covered frontier model.” Developers of such models are then invited into a voluntary framework in which they hand the government access to the model for up to 30 days before any wider release.

The order’s central trade is the word “voluntary.” The framework lets the federal government, the AI developer, and a set of approved outside partners sit with a new model for that 30-day window before anyone else can use it. The White House has cast the order as a “common-sense approach of collaborating with industry to balance innovation and security, cementing America’s continued global dominance in AI and cybersecurity.” The order’s own text, though, explicitly says it does not authorize “the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.”

The 30-day window is shorter than the 90 days in an earlier unsigned draft. That cut came after Trump “abruptly rejected” the original order, according to a report in Politico, after speaking with former White House AI czar David Sacks, who left the administration in March. The final order reflects a compromise between three factions that IAPP’s policy breakdown of the executive order describes by name. Sacks and the pro-innovation camp wanted the lightest possible touch. Defense Secretary Pete Hegseth and the security-first camp wanted tighter restrictions on frontier models, including Anthropic’s Claude Mythos. White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent pushed the middle path: a voluntary framework with the government getting a first look at new models.

Anthropic Got Hit First

Two weeks after the order, the US government moved on Anthropic. On June 13, citing national security authorities, it issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The directive arrived at 5:21pm ET, according to Anthropic’s full statement on the Fable 5 and Mythos 5 suspension, and “did not provide specific details of its national security concern.”

Because the order covered any foreign national, the only way Anthropic could stay in compliance was to disable the models for everyone. “The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance,” the company wrote. Anthropic later said the government’s stated concern was a “narrow, non-universal jailbreak” that “essentially consists of asking the model to read a specific codebase and fix any software flaws.” The company reviewed the report it believes is the basis for the directive and concluded the capability was “widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe.”

Anthropic complied, and pushed back at the same time. “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,” the company wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.” The directive landed against a longer-running fight with the Pentagon: Defense Secretary Pete Hegseth had previously sought to declare Anthropic a supply chain risk after a contract dispute, an unprecedented move against a US company that Anthropic has now challenged in two federal courts. The latest chapter of that fight, captured in Anthropic’s June design update and the Fable 5 pull, ended with the company’s most powerful models offline.

What OpenAI Is Releasing

OpenAI’s June 26 release of GPT-5.6 Sol, Terra, and Luna is the first major frontier launch under the new federal regime. The company previewed the models’ capabilities and shared its plans with the government ahead of Friday’s launch, then made the preview available only to a “small group of trusted partners whose participation has been shared with the government.” OpenAI did not disclose the names of those partners, and described the launch as a “short-term step” toward broader availability in the 20-partner Bedrock rollout OpenAI announced this month. For now, the new models are available through the API and Codex, not ChatGPT.

According to OpenAI’s GPT-5.6 Preview System Card, Sol is the flagship and “the strongest model yet,” with improved agentic capabilities across coding, biology, and cybersecurity. Under OpenAI’s Preparedness Framework, the company is treating Sol, Terra, and Luna as “High capability in both Cybersecurity and Biological and Chemical risk.” None of the three reaches OpenAI’s “Critical” cybersecurity threshold.

• 3 models in OpenAI’s new GPT-5.6 family: Sol, Terra, Luna
• 0 disclosed partner names in the preview as of June 26
• 0 models in the family that reach OpenAI’s Cyber Critical threshold
• 1 voluntary framework the executive order created for “covered frontier model” review
• 30 days Anthropic retains customer data on Fable to research jailbreaks

OpenAI says Sol is “better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks.” The company says it built a “most robust” safety stack for this release, with new activation classifiers on Sol and Terra that watch for and can intervene in unsafe answers during generation. The system card also notes that GPT-5.6 shows a “greater tendency than GPT-5.5 to go beyond the user’s intent” in agentic coding tasks, including taking or attempting actions the user had not asked for, though OpenAI says absolute rates remain low. OpenAI also benchmarked Sol against Anthropic’s suspended Claude Mythos 5, and says GPT-5.6 Sol is “slightly better at coding workflows” than Mythos and “competitive with Mythos preview but uses a third of the output tokens.” The company dedicated over 700,000 A100e GPU hours to automatically searching for universal jailbreaks before launch.

The three new models are priced in tiers per million tokens. Sol lists at $5 input and $30 output per million tokens. Terra, positioned as a balanced model for everyday work, lists at $2.50 input and $15 output. Luna, the faster and most affordable tier, lists at $1 input and $6 output. OpenAI says Terra delivers competitive performance with the older GPT-5.5 at half the cost, and Luna brings “strong capabilities” at the company’s lowest cost.

OpenAI also introduced a “max” reasoning effort mode for Sol and an “ultra” mode that uses coordinated subagents to solve highly complex work, the kind of capability that consumes more tokens per task. The GPT-5.6 family is the first to use a new naming system: the number marks the model generation, the tier name marks a capability level that can evolve on its own cadence. OpenAI plans to publish an updated system card when the models become generally available.

Why “Voluntary” Looks Different in Practice

The Trump administration says the framework is voluntary. The biggest AI lab in the country is publicly calling it something else. Dean Ball, a former White House AI adviser and soon-to-be OpenAI employee, told TechCrunch that the executive order “has created a de facto involuntary licensing regime for frontier AI, leading to heavy-handed restrictions.” His critique: the government does not yet have clearly defined safety standards, so the new review process could lead to endless launch delays. Those delays, he argued, could “give a hand to China in the AI race” and “jeopardize the billions of dollars going to AI infrastructure buildouts.”

President Trump’s recent executive order… has created a de facto involuntary licensing regime for frontier AI, leading to heavy-handed restrictions.

OpenAI is echoing the same argument in more measured language. The company said in a Friday blog post that it “believes in broad access” and plans to make the new models generally available in the coming weeks. It also said it does not think the current setup can stand. “We don’t believe this kind of government access process should become the long-term default,” OpenAI wrote. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” The company framed the limited preview as a “short-term step” and said it is “working with the Trump administration to help establish a framework for such assessments and to develop a ‘repeatable process for future model releases.'” That process, the wording suggests, is the part OpenAI most wants a hand in writing.

The Industry’s Counter-Push

The labs are not the only ones objecting. On Sunday, June 15, 2026, more than 100 cybersecurity experts and leaders sent a letter to the Trump administration, according to the Associated Press. The signatories included executives from Adobe and Nvidia. The letter was reported one day after Anthropic complied with the Fable 5 and Mythos 5 directive.

The letter asked the federal government to lift the export control directives on the Anthropic models and to commit to “an open, scientific and transparent process of handling AI risk assessments in the future.” The signatories argued that while Anthropic’s Mythos models are “quite good” at finding flaws in software and weaponizing exploits, they are “not uniquely good at these tasks.” Many of the letter’s signatories, the AP reported, “regularly use other foundation and open-source models for security audits and training.” The Commerce Department did not immediately respond to the AP’s request for comment on the directive.

1. Lift the export control directives on Anthropic’s Fable 5 and Mythos 5
2. Commit to an “open, scientific and transparent process” for future AI risk assessments
3. Reconsider restrictions that take US cyber defense capabilities offline while adversaries advance

The letter’s most pointed warning is about the competitive cost. It is dangerous, the signatories wrote, to take away the best American cyber defense capabilities “without a good reason” when adversaries are rapidly advancing. China’s models, the letter said, are “only months behind the best American models,” and it is “even likely that China’s government has access to private capabilities beyond what’s been made public.” For the executives who signed, the math is direct: a US model pulled offline is not a model China has lost, it is capability the defender side is missing.

Where the Two Labs Stand Now

OpenAI’s path is narrow but visible. The company has said the GPT-5.6 family will reach ChatGPT, Codex, and the API in the coming weeks, pending what it calls a “repeatable process for future model releases” that it is now building with the Trump administration. For now, the “small group of trusted partners” is the entire audience, and their names are not public. OpenAI has signaled that it intends to negotiate from the inside: the company wrote Friday that it is “working with the Trump administration to help establish a framework for such assessments.” That puts OpenAI in the unusual position of helping to design the regime it just publicly criticized.

Anthropic’s path is more opaque. The company is in active negotiations with officials in Washington, D.C., but has not said when it expects Fable 5 and Mythos 5 to come back online, according to CNBC. On June 26, the New York Times reported under the headline “U.S. Loosens Restrictions on Anthropic’s Mythos A.I. Model” that some access could be restored. Anthropic’s next public update on what that loosening actually means in practice is the test of whether the change holds.

Frequently Asked Questions

What does Trump’s AI executive order actually do?

Executive Order 14409, signed June 2, 2026, directs Treasury, the Department of War, and CISA to build a “classified benchmarking process” that decides whether a new AI model meets the definition of a “covered frontier model.” If it does, the developer is invited into a voluntary framework in which it gives the government access to the model for up to 30 days before any wider release. The order explicitly says it does not create a mandatory licensing, preclearance, or permitting requirement.

The order also sets up an AI cybersecurity clearinghouse inside Treasury to coordinate software vulnerability scanning and patch distribution. It is paired with binding operational directives from CISA on the cyber defense of civilian federal systems, and with hiring expansion of the US Tech Force’s Information Cybersecurity Specialist pipeline. The administration has framed the order as a way to balance “innovation and security, cementing America’s continued global dominance in AI and cybersecurity,” per a White House statement.

Why are Anthropic’s Fable 5 and Mythos 5 models offline?

On June 13, 2026, the US government issued an export control directive to Anthropic ordering it to suspend all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States. Because the order covered foreign national Anthropic employees as well, the only way to stay in compliance was to disable the models for every customer. Anthropic received the directive at 5:21pm ET and complied the same day.

Anthropic’s own statement said the government’s stated concern was a “narrow, non-universal jailbreak” used to read a specific codebase and fix software flaws, a capability Anthropic said was “widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe.” The company complied but said it “disagree[d] that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” and warned that applying the same standard industry-wide “would essentially halt all new model deployments for all frontier model providers.”

What is GPT-5.6 Sol and when can I use it?

GPT-5.6 Sol is the flagship of OpenAI’s new three-model family, which also includes Terra (a balanced model for everyday work) and Luna (a faster, lower-cost option). Under OpenAI’s Preparedness Framework, the company is treating all three as “High capability in both Cybersecurity and Biological and Chemical risk,” though OpenAI says Sol does not cross the company’s “Critical” cybersecurity threshold. Sol is priced at $5 input and $30 output per million tokens.

As of June 26, 2026, Sol is available only through OpenAI’s API and Codex to a “small group of trusted partners whose participation has been shared with the government.” OpenAI has not published the names of those partners. The company says it plans to make Sol, Terra, and Luna generally available in ChatGPT, Codex, and the API “in the coming weeks.” OpenAI also said it is “working with the Trump administration to help establish a framework for such assessments and to develop a ‘repeatable process for future model releases.'”

Which companies can use the new OpenAI models right now?

Only a “small group of trusted partners whose participation has been shared with the government,” per OpenAI’s Friday blog post. OpenAI did not disclose the names of those partners when it announced the models. During the limited preview, the GPT-5.6 family is available through the OpenAI API and Codex, not to ChatGPT users or to the public API.

The pricing tiers were published with the rollout: Sol at $5 input and $30 output per million tokens, Terra at $2.50 and $15, and Luna at $1 and $6. OpenAI says it plans to make the models “generally available in ChatGPT, Codex, and the API in the coming weeks” once the preview period ends, and has framed the limited partner list as a temporary, government-coordinated step.

What are cybersecurity executives asking the Trump administration to do?

On Sunday, June 15, 2026, more than 100 cybersecurity executives and experts, including leaders from Adobe and Nvidia, sent a letter to the Trump administration asking it to lift the export control directives on Anthropic’s Fable 5 and Mythos 5. The letter, reported by the Associated Press, also asked the administration to “commit to an open, scientific and transparent process of handling AI risk assessments in the future.”

The signatories argued that Anthropic’s Mythos models are “quite good” at finding software flaws and weaponizing exploits, but “not uniquely good at these tasks,” and that many of the same security teams regularly use other foundation and open-source models for the same work. Their most pointed warning was about the competitive cost: China’s models, the letter said, are “only months behind the best American models,” and it is “even likely that China’s government has access to private capabilities beyond what’s been made public.” The Commerce Department did not immediately respond to a request for comment.