- US Department of Commerce conducts additional testing under the White House’s recently established review protocol.
OpenAI is set to publicly release GPT-5.6, its most capable model to date, on Thursday, following a delay imposed last month at the request of the US government over national security concerns that powerful AI systems could be exploited by adversarial nations.
The launch marks a significant milestone in the evolving relationship between frontier AI developers and Washington, as regulators grapple with how to oversee technology that is advancing faster than the policies designed to govern it.
The GPT-5.6 rollout comes just one week after the US government lifted similar restrictions on Anthropic’s latest Fable and Mythos AI models. Those models had been grounded for less than three weeks after the company was ordered to suspend access pending a national security review.
The back-to-back clearances signal that Washington’s new oversight framework for frontier AI is moving from theory into practice — and, for now, at least, America’s leading AI labs are navigating it successfully.
Axios, which first reported the OpenAI launch, cited a source familiar with the matter who confirmed that the US Department of Commerce had approved a broad release of GPT-5.6 after conducting additional government testing under the White House’s recently established review protocol.
Washington’s growing caution
The scrutiny reflects mounting concern in Washington that advanced AI systems could be weaponised by military or intelligence agencies in China, Russia, and other nations deemed adversarial.
The fear is not merely theoretical: frontier models are increasingly capable of tasks ranging from sophisticated code generation to strategic analysis, raising the stakes for any technology that could fall into the wrong hands.
The oversight push began in earnest when President Donald Trump signed an executive order creating a voluntary framework under which AI developers offer “covered frontier models” to the US government for assessment up to 30 days before releasing them to trusted partners.
While the framework remains voluntary in name, the practical reality — as demonstrated by the delays imposed on both OpenAI and Anthropic — suggests that compliance has become a de facto requirement for any lab seeking to release models at the cutting edge.
Three models, three tiers
OpenAI confirmed the launch plans in a post on X late Tuesday, detailing a three-model lineup under the GPT-5.6 banner:
- GPT-5.6 Sol — the flagship model, representing OpenAI’s most advanced capabilities yet.
- Terra — a mid-tier alternative designed to balance performance with lower cost.
- Luna — the most cost-efficient option in the family, optimized for broader accessibility.
This tiered approach mirrors an industry-wide trend among AI developers to offer differentiated pricing and capability levels, allowing customers to match the model to the task rather than paying for peak performance when a lighter option will suffice.
The jailbreak question
Amid the regulatory maneuvering, a deeper technical concern continues to shadow the field. Anthropic has publicly warned that it is “probably impossible” to make any AI model fully robust to jailbreaks — the techniques adversaries use to bypass a model’s safety guardrails.
The company further cautioned about the potential emergence of a universal jailbreak capable of unlocking “an entire class of harmful behaviours” across frontier systems.
That warning underscores the dilemma facing regulators and developers alike: even with rigorous pre-release testing, no model can be guaranteed immune to misuse once it is out in the world.
