OpenAI has begun a limited preview of GPT-5.6, a three-model family — Sol, its new flagship; Terra, a mid-tier built for everyday work; and Luna, a fast, cheap option — and for now almost nobody can touch it. Access is restricted to roughly 20 organizations, a set of trusted partners whose names OpenAI shared with the US government, at the Administration’s request, before flipping the switch. A broader release is promised for “the coming weeks.”
That sequence — brief Washington, hand over the guest list, then ship to a vetted few — is the shape of a policy that isn’t law yet. A June 2 executive order directed the government to build a voluntary framework giving federal officials secure early access to “covered frontier models,” and the White House is in advanced talks with OpenAI, Anthropic, and Google to finalize release standards as soon as this week. Reporting pegs the review window at up to 30 days before a new model can ship widely; the order calls the framework voluntary and bars formal preclearance. GPT-5.6 is what “voluntary” looks like in practice. OpenAI already delayed this family once at the government’s request, and it arrives the same season regulators briefly pulled Anthropic’s Fable 5 before its July 1 restoration — a precedent that shows how much muscle an optional rule can carry.
For the 20 who are in, the product news is aggressive pricing. Per million tokens, Sol runs $5 in / $30 out; Terra is $2.50 / $15, roughly half the cost of the GPT-5.5 flagship it supplants; and Luna lands at $1 / $6. OpenAI says Sol will run on Cerebras hardware at up to 750 tokens a second this month. The per-token price of frontier intelligence keeps falling — what changed is who’s cleared to buy at the front of the line. A week ago the story was Anthropic moving Fable 5 off subscriptions and onto metered billing. Now it’s OpenAI cutting prices, for a waitlist Washington helped vet.
Our take: The capability race has a new bottleneck, and it isn’t the model. For two years the question was who could build the smartest system; now the gating factor is who’s cleared to ship it, and when. A federal review clock and a government-vetted preview list turn “launch day” into a scheduling decision made partly in Washington — the same month OpenAI floated handing the government a 5% stake. Call the framework voluntary all you like; when the alternative to cooperating is a suspension, the incentive does the enforcing. The number that matters isn’t the price of a Luna token. It’s whether “about 20 partners” becomes “everyone” on the promised timeline — or whether the front of the line stays permanently short.
What to watch
- The fine print. If the White House formalizes its release standards this week, the length of the review clock and the definition of a “high-risk” model become the most important numbers in AI. Vague benchmarks hand regulators a discretionary veto.
- Who’s on the list. OpenAI shared about 20 names with the government but hasn’t published them. The mix of labs, enterprises, and agencies will tell you whether this is a safety review or an early-access perk.
- The widening. “Coming weeks” is doing heavy lifting. A fast, broad rollout means the gate is a speed bump; a slow one means frontier access is now rationed.
- Reciprocity. The real test is the next Anthropic or Google flagship. If they route through the same preview-then-ship gate, this is the new default — not an OpenAI quirk.