OpenAI Curbs GPT-5.6 Deployment Following Government Appeal, Advocates Against Standard Restrictions
OpenAI has announced a restricted launch of its latest AI models, limiting access to a select group of trusted partners due to a directive from the U.S. government.
The new GPT-5.6 series includes three models: Sol, the flagship model; Terra, which is designed for everyday use; and Luna, a cost-effective and speedy option. While Sol is recognized as the most advanced, the current release is subject to government oversight, with access granted only to partners whose involvement has been disclosed to the authorities.
This decision by the administration coincides with increasing pressure on AI firms to manage their cutting-edge technologies. After Anthropic launched its powerful model Fable 5, it was mandated to prevent any foreign access, leading to the model’s complete withdrawal from public use.
This scenario has sparked debate about the extent of government intervention in AI model distribution. Dean Ball, a former AI advisor in the White House, criticized President Trump’s recent executive order. This order requires certain AI providers to voluntarily subject their advanced models to a government review process that could last up to 30 days before public release. Ball suggests this has effectively established an involuntary licensing system for advanced AI models, resulting in strict limitations.
Ball warns that the absence of clearly defined safety standards from the government could lead to prolonged delays in launches, potentially allowing competitors like China to gain an edge in AI development, while also risking substantial investments in AI infrastructure.
Although OpenAI complied with the administration’s request, the company expressed its dissatisfaction with the situation.
In a blog post, OpenAI stated, “We do not believe this type of government access process should become the long-term standard. It restricts valuable tools from users, developers, organizations, cybersecurity professionals, and international collaborators who require them.”
OpenAI referred to the current limited access as a “temporary measure” and indicated that it aims to broaden distribution of GPT-5.6 in the upcoming weeks, as it collaborates with the administration to create new guidelines for cybersecurity and a consistent process for future model releases.
Specifications of GPT-5.6 Sol
According to OpenAI, GPT-5.6 Sol represents the pinnacle of its model capabilities, offering enhanced abilities in areas such as programming, biology, and cybersecurity. Sol includes a “max” effort mode for reasoning and an “ultra” mode that deploys coordinated subagents to tackle complex issues, which could significantly increase token usage.
OpenAI also claims GPT-5.6 performs well in various benchmarks, slightly outperforming Anthropic’s Claude Mythos 5 in coding tasks, which has also faced restrictions this month. The company asserts that GPT-5.6 Sol is competitive with the Mythos preview but operates with a third less token output.
To alleviate concerns regarding safety, OpenAI emphasizes that Sol includes its most comprehensive security measures to date. It is designed to be resilient against adversarial attacks and focuses on defensive cybersecurity rather than enabling offensive activities. Essentially, it aims to be tough to breach while aiding users in securing systems against threats.
Furthermore, OpenAI integrates its safety mechanisms into the core functions of the model rather than relying on an external filter. This approach contrasts with the issues faced by Anthropic’s Fable 5, which redirected queries on sensitive subjects to older models, leading to user frustration due to excessive false positives.
While GPT-5.6 is initially available only to a limited partner group, OpenAI plans to expand access for users of ChatGPT, Codex, and the API in the near future.
GPT-5.6 is offered in three tiered pricing models: Sol at $5 per million input tokens and $30 for output tokens; Terra at half that rate; and Luna at $1 and $6, respectively. Improvements in prompt caching have also been made to enhance efficiency and cost-effectiveness of repeated prompts.
