OpenAI restricts GPT-5.6 rollout to government-approved partners, calls arrangement unsustainable

OpenAI released its GPT-5.6 model lineup Friday to a limited group of partners whose participation was shared with the U.S. government, marking the first time the company has publicly restricted a model release at government request.

The GPT-5.6 lineup includes three models: Sol, the flagship model priced at $5 per million input tokens and $30 per million output tokens; Terra, a balanced model at $2.50/$15; and Luna, a faster, lower-cost option at $1/$6. All three models are restricted despite Sol being the only model designated as OpenAI's "most powerful."

Government pressure on AI releases intensifies

The restrictions follow broader Trump administration pressure on AI companies to limit advanced model releases. After Anthropic released Fable 5, the administration ordered the company to remove access for foreign nationals, prompting Anthropic to pull the model entirely. The administration also restricted Anthropic's Claude Mythos 5 this month.

According to Dean Ball, a former White House AI advisor joining OpenAI, President Trump's recent executive order requests that certain AI companies voluntarily submit advanced models for government review up to 30 days before release. Ball argues this has created a "de facto involuntary licensing regime" with heavy-handed restrictions.

OpenAI made its position clear: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

The company called the limited preview a "short-term step" and said GPT-5.6 will reach broader availability in coming weeks as it works with the administration on a new executive order framework for cybersecurity and a "repeatable process for future model releases."

GPT-5.6 Sol technical specifications

OpenAI claims GPT-5.6 Sol is its strongest model, with improved agentic capabilities in coding, biology, and cybersecurity. The model introduces "max" reasoning effort mode and an "ultra" mode using coordinated subagents for complex tasks.

According to OpenAI, GPT-5.6 Sol is slightly better at coding workflows than Anthropic's Claude Mythos 5 and competitive with Mythos preview while using one-third of the output tokens. Specific benchmark scores were not disclosed.

On security, OpenAI says Sol includes its "most robust security stack yet," hardened against adversarial attacks and optimized to favor defensive cybersecurity work over offensive exploits. The company says safety guardrails are built directly into the model's core behavior rather than relying on a separate filter layer—an apparent effort to avoid the routing issues that plagued Anthropic's Fable 5, which invisibly downrouted high-risk prompts to older models, causing user backlash.

OpenAI also claims improved prompt caching to reduce costs for repeated prompts.

What this means

This marks a significant shift in how frontier AI models reach the market. OpenAI's public criticism of the government process suggests even cooperative companies view the current approach as unsustainable. The lack of clearly defined safety standards, combined with case-by-case restrictions, creates uncertainty for AI companies planning major infrastructure investments. The situation also highlights the gap between voluntary executive orders and their practical enforcement—companies are technically submitting models "voluntarily," but face unclear consequences for non-compliance. How the administration defines the "repeatable process" OpenAI mentioned will determine whether this becomes standard practice or a short-term regulatory experiment.