Chat GPT-5.6 Launches Under Government Watch as Washington Decides Who Gets Frontier AI

The flagship, called Sol, along with its mid-tier sibling Terra and the low-cost Luna, went out only to around 20 organizations whose names were individually cleared by the US government. A general release is promised “in the coming weeks.” For now, the most capable model OpenAI has ever built is something the public cannot use, and the people who can were chosen with Washington’s sign-off.

The launch arrives two weeks after the US government forced Anthropic to pull its two most powerful models offline worldwide. Put together, the two events describe a shift that the industry has been bracing for and that is now plainly underway: governments, starting with the United States, are moving to decide who gets access to the most capable AI systems, model by model and customer by customer.

What GPT-5.6 Sol, Terra, and Luna are

GPT-5.6 is a three-model family. OpenAI describes Sol as its strongest model yet, with gains in coding, biology, and cybersecurity, and adds a “max” reasoning setting for long problems and an “ultra” mode that spreads work across coordinated subagents. Terra performs close to last generation’s GPT-5.5 at roughly half the price, and Luna is the cheapest and fastest of the three. OpenAI has also changed how it names models: the number marks the generation, while Sol, Terra, and Luna mark durable capability tiers.

Capability is the reason for the controls. Under its Preparedness Framework, OpenAI is treating all three models as “High” risk for both cybersecurity and biological and chemical capability, though below the “Critical” line, and below “High” for self-improvement. On internal capture-the-flag security testing, Sol scored 96.7 percent, Terra 91.84 percent, and Luna 85.19 percent. OpenAI says Sol is better at helping defenders find and fix software flaws than at running a complete attack on its own, and that it did not produce a working full-chain exploit under the conditions tested. The company also benchmarked Sol against Anthropic’s Mythos Preview, reporting comparable results on one exploitation test while using about a third of the output tokens.

Pricing runs $5 per million input tokens and $30 per million output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. During the preview the models are reaching partners through the API and Codex, and OpenAI says broader availability through ChatGPT, Codex, and the API will follow.

Why the launch is gated: the cyber executive order

OpenAI did not choose the limited release on its own. The Trump administration asked the company to stagger the rollout, the first time the US government has preemptively pushed an American AI company to restrict a model before launch. According to reporting cited by The Next Web, the government approved access customer by customer during the preview.

Sam Altman explained in detail that the company will get the new model to users as soon as it can, source: X

The request is the first real test of an executive order President Trump signed on June 2, which asks AI companies to voluntarily give the government up to 30 days of pre-release access to models with advanced cyber capabilities. The order stops short of mandatory licensing. OpenAI framed its cooperation as a way to reach a wider release faster, but it used the announcement to register an objection. “We don’t believe this kind of government access process should become the long-term default,” the company wrote, arguing that gatekeeping keeps useful tools from developers, enterprises, and cyber defenders. OpenAI called the preview a short-term step while it works with the administration on a cyber executive order framework and what it described as a repeatable process for future releases.

Also, the fact that the 3 models share names with alt coins was not lost on the crypto community. Sol, Luna and Terra? Seriously?!

Solana noted the naming convention and its crypto similarities, source: X

The Anthropic precedent: Fable 5 and Mythos 5 pulled offline

The reason OpenAI’s pushback reads as cautious rather than defiant is what happened to its main rival. On June 12, the Commerce Department issued an export-control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, including the company’s own non-citizen employees. Because there is no way to screen users by nationality in real time across a base of hundreds of millions, Anthropic took both models down for everyone, worldwide, within hours. BNC covered the shutdown as it happened in Washington Pulls the Plug on Anthropic’s Most Powerful AI.

It was the first time the government used export controls on AI models themselves rather than on the chips that train them. The stated trigger was a reported method for jailbreaking Fable 5. Anthropic said it had received only verbal evidence of a narrow, non-universal bypass that amounted to asking the model to read a codebase and fix the flaws it found, and that other public models could surface the same issues without any jailbreak. The company said governments should be able to block unsafe deployments, but only through a statutory process that is “transparent, fair, clear, and grounded in technical facts,” and that this action did not meet that test. Commerce Secretary Howard Lutnick delivered the order in a letter to chief executive Dario Amodei.

The friction did not start there. The administration had already moved against Anthropic earlier in the year over a federal procurement dispute, and the company is contesting a Pentagon “supply chain risk” designation in court. Anthropic’s launch of Fable 5 and Mythos 5, and the unusual decision to ship one model as two products separated by safety classifiers, is the subject of an earlier BNC piece, Anthropic Splits the AI Frontier in Two.

What it means for access, sovereignty, and enterprises

The immediate effect lands on three groups. Enterprises that wire a single frontier model into production now have to treat sudden loss of access as a real operating risk, not a hypothetical one. Anthropic’s customers learned that a model can be switched off by a letter from Washington with no notice and no published evidence. OpenAI’s preview customers are learning that access can be granted, and presumably revoked, through a list the government helps maintain.

The second group is the rest of the world. Britain’s AI and online safety minister, Kanishka Narayan, responded to the Anthropic shutdown by arguing that access to advanced AI is now a matter of national capability and a reason for countries to invest in their own. Expect more governments to reach the same conclusion. If the most capable models can be cut off at the US border, sovereign AI stops being a slogan and becomes a procurement decision.

The third group is the labs themselves, which are testing two different playbooks. OpenAI is cooperating in public while objecting to the principle, betting that flexibility buys a faster path back to broad release. Anthropic complied with its order but called it disproportionate and warned that the same standard applied across the industry would halt frontier deployments altogether. Both companies are heading toward public listings, which makes the question of who controls their core product more than a policy abstraction.

For the past few years the working premise was that each new frontier model would be broadly available on release. After Fable, Mythos, and now GPT-5.6, the safer assumption is that the most capable models arrive gated, and that a government decides when the gate opens.