OpenAI Limits GPT-5.6 Rollout After Government Request, Expresses Discontent

OpenAI has announced a significant limitation on the release of its latest AI models, the GPT-5.6 lineup, to a "small group of trusted partners" following a direct request from the U.S. government. This restriction impacts all three models: Sol, the flagship and most powerful model; Terra, designed for balanced everyday use; and Luna, a faster, more cost-effective option. The company confirmed that the preview access is exclusively for partners whose participation has been pre-approved by the government, signaling a new era of governmental oversight in frontier AI development. This move by the Trump administration is part of a broader push to impose stricter controls on advanced AI systems. A similar incident recently saw Anthropic ordered to remove access for foreign nationals to its Fable 5 model, leading to the model's complete withdrawal. Dean Ball, a former White House AI adviser soon to join OpenAI, highlights that President Trump's recent executive order, which asks AI companies to voluntarily submit models for review 30 days prior to release, has effectively created an involuntary licensing regime. Ball warns that a lack of clear safety standards could cause endless launch delays, potentially benefiting competitors like China and jeopardizing massive investments in AI infrastructure. Despite complying with the administration's request, OpenAI has voiced its strong disapproval of the arrangement. In a recent blog post, the company stated, "We don’t believe this kind of government access process should become the long-term default." OpenAI argues that such restrictions prevent valuable tools from reaching users, developers, enterprises, cyber defenders, and global partners who critically need them. The company views this limited preview as a "short-term step" while it collaborates with the administration to establish a new executive order framework for cybersecurity and a "repeatable process for future model releases." The GPT-5.6 Sol model is touted as OpenAI's most robust offering yet, featuring enhanced agentic capabilities across coding, biology, and cybersecurity. It introduces a "max" reasoning effort mode and an "ultra" mode that utilizes coordinated subagents to tackle highly complex tasks, albeit with increased token usage. OpenAI claims Sol outperforms Anthropic's Claude Mythos 5 in coding workflows and offers competitive performance with Mythos preview while consuming a third of the output tokens. OpenAI has also emphasized the advanced safety features integrated into Sol. The model incorporates its most robust security stack to date, heavily fortified against adversarial attacks and specifically optimized to prioritize defensive cybersecurity work over offensive exploits. This means Sol is designed to be highly resistant to "jailbreaking" and focuses on educating users on defense mechanisms rather than facilitating system breaches. Crucially, OpenAI's safety guardrails are embedded directly into the core model's behavior, avoiding the pitfalls encountered by Anthropic's Fable 5, which used an external filter that led to false positives and user dissatisfaction by routing high-risk prompts to older, less capable models. While initial access is limited, OpenAI plans to make the GPT-5.6 models more widely available to users of ChatGPT, Codex, and the API in the coming weeks. The models will be offered in three sizes with tiered pricing: Sol at $5 per million input tokens and $30 per million output tokens; Terra at half that cost; and Luna at $1 and $6, respectively. OpenAI has also improved prompt caching to make repeated prompts more economical and predictable.