GPT-5.6 previewed by government permission
On 26 June 2026 OpenAI showed the GPT-5.6 generation — Sol, Terra and Luna — but opened access only to a narrow group of partners agreed with the US government, at the request of two White House offices. The company complied and publicly disagreed with the arrangement.
Why it matters
What marks this day is not speed or scores but who decided: for the first time access to a frontier American model was limited at the state's request before release. It is the voluntary framework from the June executive order tested in practice, the one a former White House adviser called de facto involuntary licensing.
The announcement came on 26 June 2026. Preview access went to a narrow group of partners whose participation, the company said, had been shared with the government; the restriction covered all three tiers of the generation and not the most powerful one alone. The stated reason was the models' cyber capabilities. The company called the step short-term and said this kind of government access process should not become the long-term default, because it keeps the best tools from users, developers, enterprises and cyber defenders. The basis for the request was the voluntary framework in the executive order of early June, under which companies may submit their most powerful models for government review up to 30 days before release. Dean Ball, a former White House AI adviser, called this a de facto involuntary licensing regime for frontier AI and pointed to the absence of clearly defined safety standards. Alongside the preview OpenAI described a max reasoning mode and an ultra mode with coordinated subagents, and said its guardrails are built into the model's own behaviour rather than laid over it as a separate filter. The general release came on 9 July; this record covers the limited preview only and asserts no broader availability on its date.