OpenAI is restricting the rollout of GPT-5.6 after government request, and says restrictions should not be the norm

OpenAI is limiting the release of its latest AI models to a “small group of trusted partners” at the request of the U.S. government, the company said on Friday.

The next-generation GPT-5.6 lineup includes the flagship Sol; Terra, a more balanced model for everyday use; and Luna, a faster and less expensive option. Although Sol is the company’s strongest position, the Trump administration has restricted the launch of all three. OpenAI said the preview is limited to partners “whose engagement has been shared with the government.”

Management request This comes as the US government puts new pressure on AI companies to restrict their more advanced systems. After Anthropic released its most powerful public model Fable 5, management ordered the company to remove access for any foreign nationals, prompting Anthropic to remove the model entirely.

The incident raised questions about how much authority the government should have over versions of AI models. Dean Paul, former White House advisor on artificial intelligence He will soon become an employee at OpenAIsays President Trump Recent executive order — which requires some AI companies to voluntarily submit their most advanced models for government review up to 30 days before release — creating The de facto involuntary licensing system For frontier artificial intelligence, leading to severe limitations.

The problem is exacerbated, Paul says, when the government doesn’t have clearly defined safety standards, which can lead to endless delays in launches, which could not only not help China’s AI race but also jeopardize billions of dollars that go into building AI infrastructure.

While OpenAI did what management asked this time, the AI ​​company made clear that it was not satisfied with the arrangement.

“We do not believe that this type of government access process should become the default in the long term,” he said in a newspaper on Friday Blog post. “It preserves the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

OpenAI described the preview as a “short-term step” that would put GPT-5.6 on the path to wider availability in the coming weeks, as the company works with the administration to develop a new executive order framework on cybersecurity, as well as an “iterative process for future model releases.”

GPT-5.6 Sol specifications

OpenAI says GPT-5.6 Sol is its most powerful model to date, with enhanced agent capabilities in programming, biology, and cybersecurity. Sol offers a “maximum” reasoning effort mode and a “super” mode that uses coordinated sub-agents to solve very complex tasks (just a kind of neat trick that increases your token usage).

OpenAI says GPT-5.6 excels on several criteria, including being slightly better at encrypting workflows than Anthropic’s Claude Mythos 5, which the Trump administration also effectively banned this month. GPT-5.6 Sol is also competitive with Mythos Preview, but uses a third of the output tokens, OpenAI says.

To allay any fears that its powerful models are insecure, OpenAI says Sol includes its strongest security suite yet. OpenAI says it is heavily hardened against adversarial attacks and is intentionally optimized to favor defensive cybersecurity actions over offensive exploits. In other words, it’s designed to be difficult to jailbreak, with a priority on showing users how to defend against exploits, rather than how to hack systems.

OpenAI also says that its safety barriers are built directly into the behavior of the underlying model, rather than relying on a separate filter on top of it. The company is likely trying to avoid the trap that Anthropic fell into with Fable 5. In the brief moments that Fable 5 was available, whenever the model’s classifiers detected a high-risk topic—such as cybersecurity, biology, or chemistry—they would not only block the claim; It will redirect the request to an older model. An overly cautious flow and blind guidance led to many false positives and user backlash.

While GPT-5.6 models are initially only available to a select group of partners, OpenAI plans to make them more widely available to people using ChatGPT, Codex, and APIs soon.

GPT-5.6 comes in three sizes with tiered pricing: Sol costs $5 per million input codes and $30 per million output codes; A terra costs half that; Luna costs $1 and $6 respectively. OpenAI says it has also improved fast caching to make recurring claims cheaper and more predictable.