OpenAI curbs GPT-5.6 rollout at government request, calls restrictions unusual

OpenAI has confirmed it is limiting the rollout of its latest GPT-5.6 model lineup to a “small group of trusted partners” following a direct request from the U.S. government. The company made the announcement on Friday, signaling an unusual level of federal intervention in the release of frontier AI technology.

The GPT-5.6 family includes three distinct models: Sol, the flagship system with the highest raw capability; Terra, a more balanced option designed for everyday applications; and Luna, a faster, lower-cost alternative. While Sol represents OpenAI’s most advanced model to date, the Trump administration’s restrictions apply to all three. According to OpenAI, the preview is confined to partners “whose participation has been shared with the government.”

This request aligns with a broader push by the administration to tighten controls on advanced AI systems. Earlier, when Anthropic released its most powerful public model, Fable 5, the government ordered the company to cut off access for foreign nationals. Anthropic responded by pulling the model entirely.

The situation has sparked a debate over the extent of government authority in regulating AI releases. Dean Ball, a former White House AI adviser who will soon join OpenAI, argues that President Trump’s recent executive order has effectively created an involuntary licensing regime for frontier AI. The order asks certain AI companies to voluntarily submit their most advanced models for government review up to 30 days before release, but Ball contends it has led to heavy-handed restrictions in practice.

The problem worsens, Ball says, when the government lacks clearly defined safety standards. Without them, launch delays could become indefinite, potentially handing an advantage to China in the AI race and jeopardizing the billions of dollars invested in AI infrastructure.

OpenAI complied with the administration’s request this time, but the company made its displeasure clear. “We don’t believe this kind of government access process should become the long-term default,” the company wrote in a Friday blog post. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

OpenAI described the current preview as a “short-term step” that should lead to broader availability within weeks. The company is working with the administration to develop a new executive order framework focused on cybersecurity, as well as a “repeatable process for future model releases.”

GPT-5.6 Sol is OpenAI’s strongest model yet, boasting improved agentic capabilities in coding, biology, and cybersecurity. It introduces a “max” reasoning effort mode and an “ultra” mode that uses coordinated subagents to tackle highly complex tasks. OpenAI says Sol performs slightly better at coding workflows than Anthropic’s Claude Mythos 5, which the administration also effectively banned this month. Sol is also competitive with Mythos preview while using only a third of the output tokens.

To address safety concerns, OpenAI says Sol includes its most robust security stack to date. The model is heavily hardened against adversarial attacks and intentionally optimized to favor defensive cybersecurity work over offensive exploits. In short, it is designed to be difficult to jailbreak and prioritizes showing users how to defend against exploits rather than how to hack systems.

OpenAI also notes that safety guardrails are built directly into the core model’s behavior, rather than relying on a separate filter on top. This approach appears designed to avoid the pitfalls that plagued Anthropic’s Fable 5. When that model was briefly available, its classifiers would route high-risk prompts to an older model instead of simply blocking them. The overly cautious flow and invisible downrouting led to many false positives and user backlash.

While the GPT-5.6 models are initially available only to a select group of partners, OpenAI plans to make them accessible to a wider audience through ChatGPT, Codex, and the API soon. The lineup comes in three sizes with tiered pricing: Sol costs $5 per million input tokens and $30 per million output tokens; Terra is half that price; and Luna costs $1 and $6, respectively. OpenAI has also improved prompt caching to make repeated prompts cheaper and more predictable.