OpenAI stages GPT 5.6 launch after US government request

OpenAI is staggering the release of its newest AI model, GPT 5.6, after a direct request from the Trump administration.

The company will give a small group of vetted US partners early access before a broader launch in the coming weeks. OpenAI said it briefed federal officials on the model's capabilities ahead of Friday's announcement, and that the staged approach came at the government's request rather than its own choice.

The arrangement makes OpenAI the second major US AI lab in a matter of weeks to have its launch schedule set by Washington rather than its own roadmap, after a near identical dispute over Anthropic's Mythos model. For an industry that has set its own pace since ChatGPT's 2022 debut, that marks a real shift in who controls the release calendar.

GPT 5.6 ships in three tiers. Sol is the most capable, Terra trades some performance for a lower cost, and Luna is the cheapest option for high volume use. OpenAI said Sol was its strongest model yet but did not cross a cyber critical threshold under its internal safety framework. The company added that Sol was better suited to helping people find and patch security flaws than to carrying out attacks unsupervised from start to finish.

That distinction matters because it is the same line Anthropic drew, and failed to hold, with Mythos. Anthropic's own safety testing initially placed Mythos below the threshold that would trigger the strictest export controls. The US government disagreed once it reviewed the model independently, and ordered restrictions anyway. OpenAI now faces the same independent review for GPT 5.6, with Sol the version most likely to draw scrutiny given its cyber capabilities.

Chief executive Sam Altman told staff this week that two federal bodies, the Office of the National Cyber Director and the Office of Science and Technology Policy, had requested the staged rollout. The government will approve access customer by customer during the preview period, with wider availability expected within weeks if that process goes smoothly. Altman said the arrangement was not the company's preferred long-term approach, and that OpenAI would keep working with officials and the wider industry to settle on something more sustainable for future model launches.

OpenAI was blunt about its discomfort with the process in a public blog post, arguing this kind of government access process should not become the long-term default. The company framed the delay as a short-term step taken while it works with the White House on a formal vetting framework, created under an executive order President Trump signed earlier this month. That framework is voluntary on paper, but neither OpenAI nor Anthropic has so far tested what happens if a lab declines to take part.

The rollout splits access along national lines:

• All organisations receiving early GPT 5.6 access will be US-based
• Foreign partners are expected to be added within a week
• Staff at those US firms based abroad in supported countries, including the UK and Australia, will retain access regardless

Commerce secretary Howard Lutnick pushed back against even this limited release, calling Altman directly to demand sign-off from additional agencies before launch. That intervention suggests the staged rollout was not a quiet formality inside the administration but a live point of disagreement between departments, even before GPT 5.6 reached its first external user.

OpenAI's position differs from Anthropic's in one important respect. Anthropic delayed Mythos voluntarily at first, but the government later went further and ordered the company to block foreign nationals from public access entirely, citing the model's cyber-hacking capabilities, a step Anthropic detailed in its dispute with the Pentagon. OpenAI's restriction so far remains a government-requested preview rather than an enforced block, though the gap between the two outcomes has narrowed twice already this year.

The shift marks a reversal from the White House's earlier deregulatory instincts. Vice-president JD Vance argued last year that excessive regulation could kill a transformative industry, a position that has hardened as models like Mythos, described by the UK's AI security body as a step up over prior cutting-edge systems, have grown more capable. Two labs now answer to the same review process within a single quarter, where a year ago neither did.

Altman has said the current arrangement is not OpenAI's preferred long-term model, even as the company continues expanding the infrastructure behind ChatGPT and its underlying compute commitments. Neither OpenAI nor Anthropic has said what happens once a model clears the staged preview. Mythos has not returned to public availability since the government's order, and OpenAI has given no date for when GPT 5.6 moves beyond customer by customer approval into a standard launch.

That gap leaves both companies in an unusual position for an industry built on rapid iteration. Model releases that once took weeks from internal testing to public launch now depend on the schedules of two federal offices that were not previously part of any AI lab's product calendar. Enterprise customers waiting on Sol for security work, and developers building on Terra and Luna for cost reasons, have no fixed timeline beyond Altman's own estimate of a couple of weeks. For now, how quickly Sol, Terra and Luna reach users outside the initial preview group is a decision sitting with federal regulators, not OpenAI.