OpenAI pauses GPT 5.6 rollout after White House request

OpenAI will limit access to its newest model, GPT 5.6, giving it only to a select group of close partners during a preview period after the Trump administration asked the company to slow the public rollout. CEO Sam Altman told staff the government would be "approving access customer by customer," and OpenAI hopes a broader release could follow a "couple of weeks later" if the limited preview goes well.

What's changing with GPT 5.6's release?

GPT 5.6 will not follow the broad public distribution pattern of some prior releases, instead being shared only with a narrow set of partners and subject to government approval for each customer. The Office of the National Cyber Director and the Office of Science and Technology Policy asked for the limited release, and company staff reportedly worked closely with those agencies on the upcoming rollout.

The preview period is being run as a controlled access program rather than an immediate public launch. Altman framed the arrangement as an approvals process handled "customer by customer" and indicated OpenAI may expand access a short time after the preview if no major issues surface.

How does this compare with Anthropic's approach?

Anthropic has already kept its most powerful model, Claude Mythos, locked behind a partner-only program called Project Glasswing, arguing the frontier cyber model is too capable for open public release. Anthropic announced that Claude Mythos would only be released to a small coterie of partners through Project Glasswing, saying the model posed risks if misused.

OpenAI's planned approach mirrors Anthropic's in practical terms: both companies will restrict early access to high-capability models. The difference lies in impetus: Anthropic positioned its choice as voluntary, while OpenAI's limited preview follows explicit requests from the Trump administration and close coordination with federal agencies.

What are the safety concerns driving this decision?

Officials and researchers worry frontier cyber tools can find and exploit software vulnerabilities far faster than human analysts, creating a significant security risk for complex software systems. The specific concern with models like Claude Mythos, and by implication GPT 5.6, is that they can identify and weaponize latent bugs that act as entry points into enterprise networks.

The wider context: generative models have demonstrated abilities to write malware and, in some cases, help automate ransomware attacks. Cybercriminals have used automated tools for a long time, and more capable LLMs increase the speed and scale at which vulnerabilities can be discovered and exploited.

Why it matters

Limiting GPT 5.6 to partner previews under government oversight signals a shift toward operational checks on frontier models before broad release. Agencies asking for customer-by-customer approvals means federal actors are exercising influence over commercial model deployments, not just offering guidance. Companies building advanced models will face pressure to balance product rollouts with coordinated risk assessments, and customers may see slower access to leading systems as a result.

This also reframes the public debate about corporate voluntarism versus regulatory influence: Anthropic chose partner-only distribution citing safety; OpenAI is doing so after direct government requests and coordination. That difference matters for how policy and industry practices will evolve around next-generation AI systems.

What to watch

Watch whether OpenAI expands access beyond the partner preview after the "couple of weeks" Altman mentioned and whether federal agencies formalize the customer-by-customer approval process. Also track whether other firms adopt similar restricted-release programs or if the executive branch turns voluntary testing requests into routine pre-release evaluations.