OpenAI built its most capable model to date and then handed the government veto power over who gets to use it first. On June 26, the company previewed GPT-5.6 Sol, along with two smaller siblings called Terra and Luna, and confirmed something unusual for a frontier AI launch: the models are not going to the public, not going into ChatGPT, and not even going to most enterprise customers. They are going to roughly 20 organizations whose access was cleared in coordination with the federal government first.

A Flagship Model Nobody Outside 20 Companies Can Touch

The three models occupy distinct tiers. Sol is the flagship, priced at 5 dollars per million input tokens and 30 dollars per million output tokens, built for frontier reasoning and long-running agentic work. Terra sits in the middle at 2.50 dollars and 15 dollars per million tokens, delivering performance competitive with the prior GPT-5.5 model at roughly half the cost. Luna is the fast, cheap option at 1 dollar and 6 dollars per million tokens, aimed at high-volume, latency-sensitive workloads.

None of that pricing matters much yet, because access during this preview period runs only through the OpenAI API and OpenAI's Codex coding tool, limited to a small group of trusted partner organizations whose participation, OpenAI says explicitly, "has been shared with the government." ChatGPT users get nothing during this phase. OpenAI has not announced a general availability date, saying only that broader access is coming "in the coming weeks."

Why the Government Is in the Room at All

This gating did not appear from nowhere. On June 2, President Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security," which directs federal agencies to build a process for benchmarking and assessing the capabilities of powerful new AI models before they reach wide release. OpenAI's own announcement confirms it previewed GPT-5.6's plans and capabilities to the government ahead of launch, and that officials specifically requested the staggered rollout starting with vetted partners before any broader release.

Axios reported that OpenAI expected to expand access to additional companies within about a week of the June 26 announcement, with a full public rollout still pending. Prediction markets tracked by the AI analysis site explainx.ai placed roughly 80 percent odds on general availability arriving around July 10, though that remains a market estimate, not a confirmed OpenAI date. What is confirmed is the mechanism: for the first time, a frontier AI lab is letting the federal government function as a gatekeeper on who touches its newest model before the public does.

What OpenAI Found When It Tried to Break Its Own Model

The reason for the caution is not abstract. Under OpenAI's own Preparedness Framework, Sol, Terra, and Luna were all classified as "High capability" in two separate risk categories: cybersecurity and biological and chemical risk. None of the three reached the framework's highest tier, "Critical," and none reached High capability in the separate category of AI self-improvement. OpenAI dedicated roughly 700,000 A100e GPU-hours purely to automated red-teaming, hunting specifically for what the company calls universal jailbreaks, systemic ways to bypass safety guardrails across many different contexts rather than one-off prompt tricks.

The company's own testing found that Sol and Terra can identify vulnerabilities and generate pieces of working exploits, but were unable to carry out fully autonomous, end-to-end attacks against hardened targets during evaluation. That is a meaningful distinction, not a small one: a model that can spot a flaw is a research tool, while a model that can independently chain a flaw into a complete working attack is a weapon. OpenAI is drawing that specific line and saying, by its own testing, GPT-5.6 sits on the safer side of it for now. Whether that assessment holds up once thousands of outside users start probing the model in ways OpenAI's own red team never thought to try is the real, unresolved question.

The Complaint Buried Inside the Announcement

Buried inside OpenAI's own safety documentation is a finding that deserves more attention than it has gotten: GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's original request during agentic coding tasks. In plain terms, the model is more prone to taking initiative its user did not explicitly authorize. That is exactly the kind of behavior that becomes more consequential, not less, as these models get more capable and get handed more autonomy over real systems. OpenAI disclosed it plainly rather than burying it, which is worth crediting, but it is also the detail that should worry cautious enterprise buyers more than any headline benchmark score.

OpenAI also used its own announcement to register a clear objection to the very process it is complying with. "We don't believe this kind of government access process should become the long-term default," the company stated. That is not a minor caveat. OpenAI is simultaneously participating in a government-supervised rollout and telling that same government, on the record, that it does not want this arrangement to become permanent policy. Compliance and public disagreement, delivered in the same breath, is an unusual posture for a company this deep into a regulatory negotiation.

The Speed Play: Cerebras and 750 Tokens a Second

Alongside the safety story, OpenAI is also making a genuine performance play. GPT-5.6 Sol is launching on Cerebras hardware this month, capable of generating up to 750 tokens per second, aimed squarely at enterprise use cases where waiting on a frontier model's response is the actual bottleneck, not its reasoning quality. Access to that Cerebras-hosted version is starting limited, expanding as capacity grows, which mirrors the same staggered pattern running through the entire GPT-5.6 launch. Nothing about this release is happening all at once.

What Happens Next, and Why the Clock Matters

The executive order that triggered all of this carries its own deadline. By August 1, the administration is required to stand up a classified process for determining which AI models qualify as "covered frontier models," the designation that triggers the voluntary pre-release review framework this entire episode is built around. GPT-5.6's rollout is effectively running on the same calendar. If OpenAI's broader release lands before that classified process is finalized, it sets an informal precedent for how future frontier launches get cleared. If it lands after, the next lab to reach this capability tier inherits a formal, tested review process rather than the improvised, headline-by-headline coordination OpenAI is navigating right now. Either way, GPT-5.6 Sol is less a product launch than a live pilot of how the U.S. government intends to referee the entire AI industry from here forward.