On Friday 26 June, OpenAI launched GPT 5.6 — three model variants named Sol, Terra, and Luna — with a sentence that has never appeared in a commercial AI launch announcement: "OpenAI is rolling out GPT 5.6 Friday, but limiting access to all three versions of the new model at the behest of the US government." Sol is explicitly OpenAI's most capable cybersecurity model to date, designed for long-horizon vulnerability research. Access is limited to a vetted customer list. General availability is planned in the coming weeks.
Also on Friday, Commerce Secretary Howard Lutnick sent a letter clearing partial restoration of Anthropic's Mythos-5 — but only to approximately one hundred vetted US organisations on the Annex A list. Fable 5 remains fully offline. The picture at end of week: OpenAI's most capable cyber AI gated to a vetted list, Anthropic's most capable cyber AI gated to a shorter vetted list, and nobody outside those programmes has access to either.
For European business leaders, this is the second new AI risk category in three weeks. Two weeks ago: sovereignty risk — a vendor's model can be switched off by its home government. This week: gating risk — the most capable models may not be available to you at all on launch day. The enforcement mechanism only holds while access stays gated; when OpenAI broadens GPT 5.6 to general availability, it will face the same nationality-filtering problem Anthropic could not solve on 12 June. One more item for your finance team: Anthropic has already changed Fable 5 pricing while the model is still offline — it now bills from usage credits at double the price of Opus 4.8. That is your preview of the invoice the day it returns.
