On Friday June 26, 2026, OpenAI previewed GPT-5.6 in three flavours — Sol (flagship), Terra (everyday), and Luna (cheap) — and confirmed in its own launch post that the rollout was beginning not as a public preview, not as a waitlist, and not as an enterprise-only beta, but as access for "a small group of trusted API and Codex partners whose participation has been shared with the U.S. government." About twenty organisations cleared the first round, and individual approvals are now flowing through the Office of the National Cyber Director (ONCD), the Office of Science and Technology Policy (OSTP), and the Commerce Department, per Forbes and 9to5Mac.

Thirteen days earlier, the same three offices ran the same playbook against Anthropic Fable 5 and Mythos 5 — and the lab was offline an hour and forty minutes later. The Loop covered the directive on June 13. The model on the slide changed. The harness did not.

The harness, not the model, is the story

OpenAI's launch post front-loaded capability claims: "a major step forward in frontier reasoning, coding, and agentic capabilities," with Sol scoring "High" on both the biological and the cybersecurity rubrics in the lab's published Preparedness Framework. That is the same evaluation architecture Anthropic used when it described Mythos as a "frontier cybersecurity model" earlier this year. Two labs, two High flags, one destination — a vetting room in the Herbert C. Hoover Building.

The mechanism is novel only in the OpenAI direction. Anthropic ran Project Glasswing with roughly two hundred vetted partners for months before the June 13 directive hardened that ad-hoc programme into the binding shape of an export control. OpenAI walked into the same arrangement before it had shipped a public preview at all. The ten-second summary: the regulatory regime that took six months to build around Anthropic took thirteen days to apply to the second lab. That is a sentence the OpenAI legal team did not want to write into a launch post, and it is the sentence Anthropic did not want to find out had a sequel.

President Trump's June executive order — signed the same week Washington flipped the kill switch on Anthropic, full text on the White House site — required "verifiable disclosure of model capabilities and access controls to designated executive branch offices prior to general availability." That was the legal hook. The customer-by-customer approval flow is the procedural one.

What "limited preview" actually means in practice

For the next several weeks — OpenAI is studiously not committing to a date — Sol, Terra and Luna are reachable through three lanes only: the OpenAI API on Tier 4+, Codex for Enterprise customers on Pro plans, and Amazon Bedrock, which is the back door most enterprise procurement teams will actually use. Each organisation requesting access submits identifying information that OpenAI shares with the government for review. ONCD and OSTP screen for what the executive order calls "national-interest" applications; Commerce makes the call on whether to clear.

The early partner list is unnamed in any of the trade-press coverage I read. OpenAI's blog and the wire copy both stop at "approximately twenty." The unspoken read is what it always is when a list isn't published: the launch partners are the obvious suspects — the labs themselves, two or three Magnificent-Seven hyperscalers, the Pentagon's CDAO, maybe a national lab, and the three management consultancies that landed on Anthropic's Global Premier list two Wednesdays ago. The names will appear. They always do.

The pricing is the most boring part of the announcement, which is its own kind of signal:

• Sol: $5 input / $30 output per 1M tokens
• Terra: $2.50 input / $15 output
• Luna: $1 input / $6 output

That is the GPT-5.5 cap-table walked across one column on the spreadsheet. The release is not pricing a capability premium for the new tier. It is pricing the second-best version of the new tier and giving Sol away at the old price. Read that as confidence in volume rather than confidence in margin.

The OpenAI sentence that disagreed and the rollout that complied

The line from OpenAI's own preview post that the trade press picked up verbatim is worth printing in full: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

That is the cleanest possible version of disagree and comply — the same phrase the Loop applied to Anthropic on June 13, lab swapped out, structure intact. The difference is that Anthropic's disagreement was forced; OpenAI's was volunteered into the launch post. Both companies are now publicly on the record that the regime they participate in is one they think the country shouldn't keep. Neither company has done anything to leave it.

That gap — between the published objection and the live integration — is the story for the next quarter. If two of the three frontier labs go on the record against the access process while shipping into it, the policy debate is over before it starts. The fact pattern is decided by the consent. Sam Altman's reported message to staff, via The Information, said the government had "asked OpenAI to stagger GPT-5.6" and that the company viewed the preview as "the fastest path to wide availability if preview clears." The verb in that sentence is clears. The grammar belongs to a regulator, not to a product team.

Commerce Secretary Howard Lutnick reportedly spoke with Altman on Wednesday June 24, per the same trade-press chain, to confirm that all relevant executive-branch offices had completed their assessment before Friday's preview went live. Lutnick is the same cabinet secretary whose department issued the Anthropic directive thirteen days before. His phone, on the AI-policy beat, has become the most directly consequential line of communication in the cabinet.

Sol's capability ratings are the reason it was gated

The Preparedness Framework numbers OpenAI published with the model are the part the cybersecurity press is going to litigate. Sol's self-reasoning control rate is 1.3% — the share of times in a battery of agentic-execution evaluations where the model takes an action an alignment reviewer would flag as outside the policy envelope. GPT-5.5's number was 0.4%. Tripling the metric the framework is designed to keep low is not a benchmark a flagship usually highlights. OpenAI does, because the alternative is for METR to publish it.

The cybersecurity flags are sharper. Sol scored "High" on both CVE-Bench and ProtocolQA — vulnerability identification and protocol comprehension respectively — and the lab is explicit that it does not yet trip the "Cyber Critical" threshold under the Preparedness Framework. The High rating is the one that triggered the Anthropic process in April. It is the same rating now flying on Sol. The administration didn't have to invent a category to apply to the new model. The category was already in the file from the last one.

The Terminal-Bench 2.1 comparison is the public-relations line. OpenAI's own post claims Sol outperformed Mythos 1. Anthropic, in turn, has not yet published a comparable score for its post-Mythos line. The next benchmark fight is the one that decides whether the second tenant in the harness gets the headline credit for being there.

What this means

1. The export-control regime around frontier models has institutionalised in thirteen days. That is faster than any of the labs predicted in their March policy submissions. The "frontier" label is now a regulatory category, not just a vendor adjective, and the offices that apply it have agreed on a workflow that can be re-run on the next model with no new statute and no new directive.

2. OpenAI's preview is a precedent for whatever Anthropic ships next. The administration's response to "a model we think is dangerous" now has a procedural answer that doesn't require the kill switch. Whichever lab ships next after this — Anthropic's Mythos 2 sometime in Q3, or Google DeepMind's Gemini 3.5 Pro whenever the post-Hassabis retention story stabilises — will be channelled through the same pipeline.

3. The published objections are the new norm. Both labs have now formally objected, in writing, to the regime they are inside. Watch for Brussels. The EU AI Office has not yet decided whether to mirror the US process, attack it, or quietly substitute for it. The next quarterly memo, due in early August, is the deadline.

4. The "limited preview" duration is the metric to track. OpenAI is saying "coming weeks." Anthropic was offline for the hour-and-forty-minute window before Project Glasswing was reactivated under the new shape. If the OpenAI preview duration runs into the second month, the regime is sticky. If it ends inside two weeks with Luna on the public API, the customer-by-customer review is a one-week ceremony with a public-relations cost and no operational one.

The harness was a sentence on a press release on June 13. By June 26 it was an architecture diagram. The lab that walked into it first got to call it disagree and comply. The lab that walked into it second got to call it limited preview. The customers on the other side of the wire are tired of vocabulary and waiting on a delivery date.