OpenAI's New GPT-5.6 Models Are Here, But the White House Decides Who Gets Them First

OpenAI just launched its next flagship AI lineup, and for once, the company isn't the one deciding who gets to use it.

The ChatGPT maker rolled out three new models on Friday- GPT-5.6 Sol, Terra, and Luna, but confirmed that access is being capped at the request of the U.S. government. Instead of opening the models up to the public or even its usual developer base, OpenAI is starting with a tightly controlled preview limited to roughly 20 companies whose participation has already been cleared by federal officials.

For the first time, Washington is treating frontier AI releases less like product launches and more like exports that need a government sign-off before they reach the wider world. That's a meaningful shift in how the most powerful AI systems coming out of Silicon Valley get into people's hands.

It also signals that OpenAI isn't getting special treatment. Just weeks ago, Anthropic found itself in a nearly identical position when federal restrictions forced it to pull public access to its own top-tier models, Fable 5 and Mythos 5. OpenAI's situation makes clear this isn't a one-company problem,  it's becoming the new normal for whoever sits at the front of the AI race.

What's Actually in GPT-5.6

OpenAI is breaking from its old naming habits with this release. Rather than just bumping a version number, the company introduced three permanent capability tiers that it says will keep evolving on their own schedules going forward.

Sol is the flagship, built for the heaviest lifting- complex coding, in-depth research, and demanding agent-style tasks. Terra sits in the middle, aiming to give businesses strong performance without the cost of running the top-end model constantly. Luna is the budget-friendly option, tuned for speed and everyday tasks where raw power matters less than responsiveness.

Two new features come bundled with the lineup. A "max" reasoning setting gives Sol more room to work through especially tough problems, while a new "ultra" mode lets the model split a task across multiple coordinated sub-agents working in parallel essentially handing off pieces of a complex job to a small team of AI helpers instead of tackling it alone.

OpenAI is pricing the family on a sliding scale: Sol runs $5 per million input tokens and $30 per million output tokens, Terra is roughly half that, and Luna comes in at $1 and $6 respectively. The company also overhauled how prompt caching works, promising more predictable costs for developers who repeatedly send similar requests.

The Government's Role in the Rollout

Here's where things get unusual. OpenAI says it spent the past month walking federal officials through GPT-5.6's capabilities, including meetings CEO Sam Altman held at the White House in early June. The company expected some pushback, maybe a staggered rollout. What it didn't expect was a hard cap of around 20 partners, each individually approved by the government before getting access.

OpenAI says it plans to add more companies to that list next week, with a goal of full availability in the coming weeks. According to the company, federal officials are aware of and supportive of that broader timeline, assuming nothing concerning turns up during the additional testing window.

Still, OpenAI didn't pretend to love the arrangement. In its announcement, the company pushed back fairly directly on the idea that this should become standard practice, arguing that gatekeeping access this way ends up hurting the very people who'd benefit most, developers, businesses, cybersecurity teams, and international partners.

The company framed its cooperation as a temporary bridge rather than an endorsement, saying it's working with the administration to build out a clearer, repeatable review process tied to a forthcoming cybersecurity executive order rather than relying on ad hoc restrictions every time a new model ships.

The Cybersecurity Concern Driving All of This

The core worry from regulators isn't about chatbots writing essays. It's about how good these models have gotten at hacking.

OpenAI itself acknowledged that GPT-5.6 Sol is unusually capable when it comes to finding and patching security vulnerabilities. The company argues the model is better suited to helping defenders close security gaps than it is at independently pulling off a full cyberattack from start to finish, and says Sol's abilities still fall short of what the company defines as a "critical" risk threshold in its internal safety framework.

OpenAI says it's built its strongest safety measures yet around the new models, with the goal of supporting legitimate security work while making it considerably harder for anyone to misuse the technology for offensive hacking.

What Comes Next

The bigger story here is the framework taking shape behind all of this. Under a recent executive order, federal agencies have until August to stand up a classified process specifically for evaluating how risky new AI models are from a cybersecurity standpoint. Models that cross a certain threshold would be labeled "covered frontier models",  a designation that would trigger extra scrutiny before they're allowed to reach the public.

In other words, what's happening with GPT-5.6 right now may be a preview of how every major AI release gets handled going forward, at least for the most advanced systems coming out of U.S. labs. Whether that turns into a quick, predictable process or a recurring bottleneck for the entire industry is the question both OpenAI and Anthropic are still waiting on Washington to answer.

Read More: OpenAI Unveils Its First AI Chip 'Jalapeño', Takes Direct Aim at Nvidia