OpenAI said Friday it will introduce a preview release of its GPT-5.6 artificial intelligence model to a narrowly selected group of partners before making the model more widely available in the coming weeks. The company characterized the initial exposure as limited and said the decision to stagger access followed a request from the U.S. administration.
In a company blog post, OpenAI said it is providing the preview only to trusted partners whose inclusion has received approval from U.S. government authorities. The firm framed the arrangement as a short-term, targeted step taken at the administrations request to manage the models introduction.
OpenAI also signaled reservations about making this type of government access process a permanent feature of model rollouts. The company said the restricted release prevents a range of users - including individual developers, enterprises, cyber defenders and international collaborators - from accessing advanced tools they might otherwise need. Still, OpenAI described the temporary restriction as the best route to broader availability within weeks.
The initial preview cohort will include 20 partners. OpenAI said one of the designated access paths will be through Amazon.com Inc.s Bedrock software platform.
The administrations involvement in the rollout occurs amid escalating pressure from the White House on developers of powerful AI systems. OpenAIs announcement follows a related regulatory intervention affecting another major AI firm. Anthropic PBC suspended its most capable models two weeks ago after government directives required the company to bar foreign nationals, both within and outside the U.S., from using those models on national security grounds.
Within the GPT-5.6 family, OpenAI said the most capable variant, named Sol, is engineered to perform tasks in coding, biology and cybersecurity autonomously. Aware of heightened risks tied to such functionality, OpenAI said it has tightened safeguards for higher-risk uses of the top-tier model, particularly for sensitive cyber-related requests.
At the same time, the company acknowledged limits to pre-release evaluations. OpenAI noted that no single evaluation can capture every possible product configuration, multi-step attack scenario or real-world workflow. To manage emergent threats, the company says it maintains a rapid-response system to reproduce and assess newly discovered jailbreaks, prioritize fixes and fold those cases into ongoing evaluations to reduce the likelihood of repeat failures.
OpenAI said it hopes a recent executive order signed by President Donald Trump will help clarify future release protocols. The directive gives the administration and AI companies 60 days from signing to craft a voluntary framework. Part of that framework could permit the government to access so-called frontier models for up to 30 days prior to planned public release.
For now, OpenAIs approach represents a calibrated, short-term compromise between accelerating availability and addressing immediate national security and safety concerns. How that compromise shapes broader access policies and the pace of model deployments remains subject to the framework the administration and industry develop in the coming weeks.