OpenAI slows down GPT-5.6 rollout over government security concerns

OpenAI was ready to pull back the curtain on its next-generation models, but the White House just stepped in front of the stage.

On Wednesday, OpenAI CEO Sam Altman told employees that the upcoming GPT-5.6 rollout will be limited to a small group of enterprise partners after federal officials requested a staggered launch. The pressure came directly from the Office of the National Cyber Director and the Office of Science and Technology Policy, with Commerce Secretary Howard Lutnick separately warning OpenAI against launching without broader interagency greenlights.

Altman said he hopes a full public launch will follow a couple of weeks later, but for now, access requires government awareness, and reportedly, case-by-case federal approval.

Sol gets benched alongside the whole family

The Trump administration did not stop at OpenAI's most capable model. The restriction covers the entire GPT-5.6 lineup: the flagship Sol, the balanced Terra, and the lightweight Luna.

According to OpenAI, Sol is its most powerful system to date. It introduces a max reasoning mode and an ultra mode that spins up coordinated subagents to tackle massive tasks. OpenAI claims Sol edges out Anthropic's Claude Mythos 5 in coding benchmarks while burning just one-third of the output tokens. On the pricing sheet, Sol sits at $5 per million input tokens and $30 per million output tokens, with Terra at half price and Luna at $1 and $6.

OpenAI also went out of its way to stress its defense-first safety architecture. Unlike Anthropic's short-lived Fable 5, which routed sensitive prompts down to older models and triggered massive user backlash, OpenAI says Sol bakes safety straight into its core weights, explicitly favoring defensive cybersecurity over offensive exploits.

None of that was enough to avoid a delay. In a Friday post, OpenAI pushed back unusually hard, stating that this kind of government gatekeeping should not become the long-term default because it keeps critical tools away from defenders and businesses.

The voluntary review is not voluntary anymore

Washington's move confirms what many AI labs feared after Anthropic's ordeal earlier this month. When Anthropic released Fable 5, the administration slapped down an export control directive barring foreign nationals, including Anthropic's own non-US staff, from using it. Anthropic had to pull the model down entirely.

While OpenAI got off slightly easier with a staggered partner preview, former White House AI adviser Dean Ball warned that the administration's recent executive order has morphed into a de facto licensing regime. Ball argued that because officials lack objective, transparent safety metrics, voluntary 30-day previews threaten to become indefinite release delays that imperil billions of dollars in infrastructure bets.

Why it matters

If you were planning to build on GPT-5.6 next week, you are out of luck. Sol's agentic coordination and aggressive token efficiency could dramatically lower the cost of complex coding and research workflows, but developers will have to wait out Washington's approval cycle. The preview list is tiny, and early access remains strictly gated.

For the broader AI ecosystem, this marks a jarring pivot. Despite campaign rhetoric promising deregulation and rapid export dominance, the administration is now wielding heavy-handed, behind-the-scenes vetos over individual model drops. That inconsistency makes compliance a moving target for every frontier lab.

We think OpenAI played its cards as well as it could by offering an olive branch preview rather than facing Anthropic's total shutdown. But it looks increasingly clear that the era of shipping frontier models on your own timetable is officially over.