OpenAI confirmed on July 8 that it will publicly release GPT-5.6, which it calls its most capable model to date, on Thursday. The launch comes about a month later than planned: the US government asked the company to hold the model back for additional security review over national security concerns, according to Reuters.
Why Washington intervened
Reuters reported that US officials worried that highly capable frontier models could end up in the hands of adversary states or other hostile actors and be used for military, intelligence or cyber operations. In recent months the administration has tightened oversight of frontier AI releases as a result.
The policy follows an executive order from President Donald Trump that encourages companies building frontier AI to voluntarily submit their most advanced models to the government for evaluation before public release. The stated aim is to identify risks to national security or global stability before a model reaches the public.
According to the report, the Commerce Department approved a broad release of GPT-5.6 under the new framework once the extra testing was complete. During the delay OpenAI had given a limited group of trusted partners access to the model to evaluate its performance and safety in real conditions.
Anthropic went through the same process
The move follows a similar episode a week earlier, when the government lifted a roughly three week hold on Anthropic’s Fable and Mythos models. Those models had also been paused on national security grounds and were cleared for global release after additional testing.
Anthropic has acknowledged that it may be impossible to make any AI model fully resistant to jailbreak attacks, the technique of crafting prompts that bypass a model’s safety restrictions. The company has said a so called universal jailbreak that defeats multiple safeguards at once could emerge, which is why safety systems are being strengthened alongside each new model generation.
Three versions of GPT-5.6
OpenAI is releasing three variants. GPT-5.6 Sol is the most capable model, Terra is a lower priced mid tier option and Luna is the lowest cost version, giving consumers and businesses a choice based on their needs and budgets.
Government security review before the release of advanced AI models looks set to become a routine step, at least in the United States, as regulators and developers work out how to handle systems with potential dual use capabilities.
