OpenAI is rolling out its latest AI models to only a “small group of trusted partners” after the U.S. government asked it to limit access. The company announced the restricted preview on Friday, saying participation is limited to partners “whose participation has been shared with the government.”
The new lineup is called GPT-5.6 and includes three models: Sol, the flagship; Terra, a balanced option for everyday use; and Luna, a faster and cheaper alternative. Despite Sol being OpenAI’s most capable model to date, the Trump administration has put a hold on all three, restricting their public release until a broader framework is in place.
OpenAI went along with the request but didn’t hide its frustration. “We don’t believe this kind of government access process should become the long-term default,” the company wrote in a blog post. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”
This isn’t an isolated case. The administration recently ordered Anthropic to pull its most powerful public model, Fable 5, after it launched, demanding that access be removed for any foreign national. Anthropic ended up taking the model down entirely. That incident put a spotlight on a growing tension between U.S. AI companies and the current administration over who gets to decide when and how powerful AI systems reach the public.
Dean Ball, a former White House AI adviser who is set to join OpenAI, has been vocal about the risks of this approach. He argues that a recent executive order asking AI companies to voluntarily submit advanced models for government review up to 30 days before release has effectively created an involuntary licensing system. His concern is that without clearly defined safety standards, the process could lead to open-ended launch delays that hand an advantage to China in the AI race and threaten the billions of dollars already being poured into AI infrastructure.
OpenAI is treating the current situation as a short-term step, saying GPT-5.6 will move toward broader availability in the coming weeks. The company says it’s working with the administration on a new executive order framework covering cybersecurity, along with a “repeatable process for future model releases.” The hope is that future launches won’t require the same ad hoc negotiations.
On the technical side, GPT-5.6 Sol is OpenAI’s strongest model yet. It brings improved capabilities in coding, biology, and cybersecurity, and introduces two new reasoning modes: a “max” mode and an “ultra” mode that uses coordinated subagents to tackle highly complex tasks. That second mode is notably token-intensive, which means costs can climb fast for heavy users.
OpenAI claims Sol edges out Anthropic’s Claude Mythos 5 on coding benchmarks, which is notable given that the Trump administration also effectively banned Mythos 5 this month. Sol is also described as competitive with a Mythos preview model while using roughly a third of the output tokens, which translates to meaningfully lower costs at scale.
Security is a central theme with this release. OpenAI says Sol is its most hardened model to date, built to resist adversarial attacks and designed to prioritize defensive cybersecurity work over offensive exploits. Crucially, the safety guardrails are built directly into the model’s core behavior rather than sitting as a separate filter on top of it.
That last point appears to be a direct response to one of the problems that caught Anthropic out. When Fable 5 was briefly available, its classifier system would silently route high-risk queries to an older, less capable model instead of simply blocking them. That invisible downrouting triggered a wave of false positives and user complaints. OpenAI seems to be trying to avoid the same trap by baking safety into the model itself rather than layering it on after the fact.
Pricing for GPT-5.6 breaks down as follows:
- Sol: $5 per million input tokens, $30 per million output tokens
- Terra: Half the price of Sol
- Luna: $1 per million input tokens, $6 per million output tokens
OpenAI has also improved prompt caching across the lineup, making repeated prompts cheaper and more predictable in cost. The models will be available through ChatGPT, Codex, and the API once the broader rollout begins.
