OpenAI ships GPT-5.6 — but only to ~20 customers the government approves one by one
\nThe export-control net just widened from one lab to two. OpenAI released GPT-5.6 on Jun 25–26 — a three-model family (Sol, its strongest model yet; Terra, GPT-5.5-class quality at ~2× cheaper; Luna, its cheapest) — but only in a limited preview after the White House asked OpenAI to restrict initial access to roughly 20 government-vetted partners, with CEO Sam Altman telling staff the administration would approve access \"customer by customer\" during the preview (MacRumors, CNN, Engadget). The request came from the Office of the National Cyber Director and OSTP; Commerce Secretary Lutnick reportedly advised against launching without cross-agency sign-off — the exact playbook used two weeks ago to dark Anthropic's Fable 5 / Mythos 5. Sol leads on agentic coding, biology and cyber, adds \"max\"/\"ultra\" reasoning modes, and is the first OpenAI model trained with a redesigned pipeline that audits for the cross-persona reward leakage blamed for GPT-5.5's alignment failure. Broader rollout is floated \"a couple of weeks\" out, contingent on the customer-by-customer process.
\nSo what: the question is no longer \"which lab gets restricted\" — it's that frontier launches now ship through a federal approval gate. If you were counting on GPT-5.6 as your Fable/Mythos fallback, your access is now a government decision; keep an open-weight Plan C live and assume gated, US-first availability.
\n ARC: MODEL-WAVE · D17CONFIDENCE: 2x SRC\n \n \nWhat shipped. After a June release window that looked lapsed (Polymarket's Jun 22–28 odds had collapsed ~83%→~18%), OpenAI quietly launched GPT-5.6 in a limited preview on Jun 25–26: Sol (flagship — \"most robust safety stack to date,\" agentic gains in coding/biology/cybersecurity, new \"max\" reasoning effort and an \"ultra\" sub-agent mode), Terra (GPT-5.5-class quality at ~2× lower cost) and Luna (lowest price). Leaks long pointed to a ~1.5M-token context; the headline is governance, not specs (MacRumors).
\nWhy it's the lead. This is the decisive event that escalates the whole export-control arc: a second US lab now ships its best model through a government approval gate. The White House — via the Office of the National Cyber Director and OSTP — asked OpenAI to limit GPT-5.6 to ~20 vetted partners because of its capabilities; Altman framed the approvals as happening \"customer by customer\" (CNN). It directly follows the Anthropic precedent and turns a one-off into a pattern (see Developing #2).
\nThe caveats. \"~20 partners\" and the \"customer by customer\" line are reported from internal remarks; OpenAI has not published a model card or pricing page for the gated tier as of this edition. The alignment-pipeline claim (cross-persona reward-leakage audit) is OpenAI's own framing of a GPT-5.5 fix, not independently verified. The \"couple of weeks\" broader-rollout timeline is contingent and unconfirmed.
\nWatch next. Whether a public model card and pricing appear; how many customers clear the gate and how fast; whether Google (Gemini 3.5 Pro, slipping to July) faces the same ask; and whether Congress folds GPT-5.6 into its open questions about the Anthropic action.
\n