OpenAI announced Friday that it will limit the release of its latest AI models to a “small group of trusted partners” at the request of the U.S. government.
The next generation GPT-5.6 lineup includes the flagship model Sol. Terra, a more balanced model for everyday use. Luna is a faster, lower cost option. Sol is the company’s most powerful mode, but the Trump administration has restricted the release of all three. OpenAI said the preview will be limited to partners that “share participation with governments.”
The administration’s request comes amid renewed pressure from the U.S. government to limit cutting-edge systems on AI companies. After Anthropic released its most powerful public model, Fable 5, the government ordered the company to remove foreign access, leading Anthropic to remove the model entirely.
The incident raised questions about how much power the government should have over the release of AI models. Dean Ball, a former White House AI advisor and soon-to-be OpenAI employee, says President Trump’s recent executive order requiring some AI companies to voluntarily submit their latest models for government review 30 days before release creates a de facto involuntary licensing system for frontier AI, leading to heavy-handed restrictions.
Ball argues that the problem will be exacerbated if governments do not have clearly defined safety standards, which could lead to endless launch delays, handing China a hand in the AI race as well as jeopardizing billions of dollars invested in AI infrastructure.
Now, OpenAI has complied with the administration’s demands, but the AI company has made it clear that it is not happy with the deal.
“We do not believe this type of government access process should become the long-term default,” Friday’s blog post said. “We secure the best tools from the users, developers, enterprises, cyber defenders, and global partners who need them.”
OpenAI said the company is working with the administration to develop a new executive order framework for cybersecurity and a “repeatable process for future model releases,” and called this preview a “short-term step” that will put GPT-5.6 on the path to broader availability in the coming weeks.
GPT-5.6 Sol specs
OpenAI says GPT-5.6 Sol is its most powerful model to date, with improved agent capabilities in coding, biology, and cybersecurity. Sol introduces a “maximum” inference effort mode and an “ultra” mode that uses tuned subagents to solve extremely complex tasks (a kind of neat trick that spikes token usage).
OpenAI says GPT-5.6 outperforms several benchmarks, including being slightly better at coding workflows than Anthropic’s Claude Mythos 5, which was also effectively banned by the Trump administration this month. GPT-5.6 Sol also competes with Mythos Preview, but uses one-third of the output tokens, according to OpenAI.
To allay concerns that the powerful model is insecure, OpenAI says Sol includes its most robust security stack to date. OpenAI says it is significantly hardened against adversarial attacks and intentionally optimized to prioritize defensive cybersecurity efforts over offensive exploits. In other words, it is designed to be difficult to jailbreak while prioritizing showing users how to protect against exploits rather than how to hack into their systems.
OpenAI also says its safety guardrails are built directly into the core model’s behavior, rather than relying on a separate filter on top of the core model. The company is likely trying to avoid the trap that caught Anthropic with Fable 5. In the brief moment that Fable 5 was available, when the model’s classifier detected a high-risk topic like cybersecurity, biology, or chemistry, it would simply block the prompt; Requests will be routed to the old model. The overall over-cautious flow and invisible downrouting resulted in many false positives and user backlash.
The GPT-5.6 model will initially be available only to a select group of partners, but OpenAI plans to make it more widely available to those using ChatGPT, Codex, and APIs in the near future.
GPT-5.6 is available in three sizes with tiered pricing. Sol costs $5 per million input tokens and $30 per million output tokens. Terra costs half that. Luna prices are $1 and $6, respectively. OpenAI says it has also improved prompt caching, making repeated prompts cheaper and more predictable.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
