Government Slams Brakes on OpenAI's GPT-5.6 Rollout Amidst Regulatory Push

OpenAI has announced a restricted release of its latest generation of AI models, GPT-5.6, to a select group of "trusted partners" at the explicit request of the U.S. government. This new lineup includes Sol, designated as the flagship and most powerful model; Terra, a more balanced option suited for everyday tasks; and Luna, designed for faster, lower-cost applications. Despite Sol being OpenAI's most advanced offering, all three models have faced restrictions from the Trump administration. OpenAI clarified that participation in this limited preview is contingent on partners whose involvement has been shared with the government, underscoring the significant government oversight.

This move by the administration reflects increasing pressure on AI companies to regulate their most advanced systems. A prior incident involving Anthropic, which was compelled to remove access to its powerful public model Fable 5 for all foreign nationals, leading to its complete withdrawal, highlights the extent of government intervention. Such events have sparked critical debates regarding the appropriate scope of governmental authority over the release of cutting-edge AI models. Dean Ball, a former White House AI adviser soon to join OpenAI, posits that President Trump's recent executive order—which mandates certain AI companies voluntarily submit their advanced models for government review up to 30 days pre-release—effectively establishes an involuntary licensing regime for frontier AI. Ball warns that this could lead to heavy-handed restrictions and, compounded by the lack of clearly defined government safety standards, might result in endless launch delays. Such delays, he argues, could not only give China an advantage in the global AI race but also jeopardize the substantial investments poured into AI infrastructure development.

While OpenAI complied with the administration's demands for this initial release, the company expressed clear reservations about the long-term implications of such an arrangement. In a Friday blog post, OpenAI stated, "We don’t believe this kind of government access process should become the long-term default," asserting that it hinders "users, developers, enterprises, cyber defenders, and global partners who need them" from accessing valuable tools. OpenAI characterized this limited preview as a "short-term step" intended to pave the way for broader availability of GPT-5.6 in the coming weeks. The company is actively collaborating with the administration to formulate a new executive order framework focused on cybersecurity and to establish a consistent, repeatable process for future model releases.

Delving into the technical specifications, OpenAI hails GPT-5.6 Sol as its most potent model to date, showcasing enhanced agentic capabilities across critical domains such as coding, biology, and cybersecurity. Sol introduces two innovative reasoning modes: a "max" effort mode and an "ultra" mode, which leverages coordinated subagents to tackle highly complex tasks, though this comes with a significant increase in token usage. OpenAI reports that GPT-5.6 excels in various benchmarks, even surpassing Anthropic’s Claude Mythos 5 in coding workflows, a model that the Trump administration also effectively banned recently. Furthermore, GPT-5.6 Sol is presented as competitive with Mythos preview while impressively using only a third of the output tokens, highlighting its efficiency.

To address potential concerns regarding the safety of its powerful models, OpenAI has equipped Sol with its most robust security stack to date. The model is described as heavily hardened against adversarial attacks and intentionally optimized to prioritize defensive cybersecurity work over offensive exploits. This means Sol is designed to be highly resistant to jailbreaking attempts and to guide users in defending against vulnerabilities rather than exploiting them. Crucially, OpenAI emphasizes that Sol's safety guardrails are intrinsically built into the core model's behavior, rather than relying on a separate, superficial filter. This approach aims to circumvent the pitfalls experienced by Anthropic with Fable 5, where a separate classifier for high-risk topics led to frequent false positives and user dissatisfaction due to automatic rerouting to older, less capable models.

Although the GPT-5.6 models are initially restricted to a select cohort of partners, OpenAI plans to make them widely accessible to users of ChatGPT, Codex, and the API in the near future. The models will be available in three distinct sizes with tiered pricing: Sol will cost $5 per million input tokens and $30 per million output tokens; Terra's pricing will be half of Sol's; and Luna will be the most economical option at $1 per million input tokens and $6 per million output tokens. OpenAI has also implemented improvements to prompt caching, designed to make repeated prompts both cheaper and more predictable for users.