OpenAI Releases GPT-5.6 to the Public After Trump Administration Security Review

OpenAI announced on Tuesday evening that it will release its entire GPT-5.6 model family, Sol, Terra and Luna, to the public on Thursday, July 10. It marks the end of a several-week period in which the most powerful version of the model, Sol, was available only to a narrow group of partners approved directly by the Trump administration.

The restriction stemmed from President Trump's June executive order, which called on developers of the most powerful AI models to voluntarily submit them for government review before launch. In practice, the situation around GPT-5.6 went further than the voluntary framework envisioned in the order, the access list was curated by the administration, not merely consulted on.

What triggered the scrutiny

The extra caution stemmed from concerns about models capable of supporting large-scale cyberattacks. In the background were references to Mythos, described as a powerful Anthropic model focused on cybersecurity that had previously raised similar concerns about potential offensive use against digital networks. OpenAI stresses that Sol is better at helping users patch vulnerabilities than at carrying out full attacks from start to finish.

After additional testing and direct talks with officials, the administration agreed to a broad rollout. OpenAI made clear, however, that it does not want this arrangement to become the standard going forward.

We don't believe this kind of government access process should become a long-term standard - OpenAI, statement at the GPT-5.6 launch

Three models, three budgets

The GPT-5.6 family splits into three tiers suited to different use cases and budgets. Sol is the flagship for tasks requiring the highest precision, particularly in the hard sciences and security. Terra is the mid-tier variant, cheaper to run while maintaining solid performance. Luna is the most economical and is aimed at mass-market applications where the cost per query matters most.

The models are said by OpenAI to include safeguards capable of withstanding real adversarial pressure, meaning attempts by users to bypass filters and extract content from the model that is dangerous or prohibited under the company's policies.

A precedent for the industry

The GPT-5.6 situation matters because, for the first time so openly, the U.S. government decided who would get access to a commercial language model before it reached the market. That goes beyond the industry's usual practice, in which safety evaluations are mostly conducted by the companies themselves, which then publish their own reports. Trump's June executive order was formally about voluntary reporting of models to regulators, but in practice it turned into a binding access list controlled by the administration.

For European and Polish companies using the OpenAI API, this introduces a new variable into deployment planning, the availability of the most powerful models may in the future depend not just on the provider's technical readiness but also on the security policy of the country where they are built. That raises questions about whether similar review mechanisms will appear for other labs, such as Anthropic or Google DeepMind, and how it will affect the pace at which new models reach customers outside the U.S.

What's next

At the same time, OpenAI is expanding access to an earlier testing stage of GPT-5.6 worldwide, and alongside the text model family it also announced a new generation of voice models, GPT-Live, designed to listen and speak simultaneously in real time. The week is also marked by rival launches, Meta rolled out its first image-generation model, Muse Image, while xAI, under its new SpaceXAI brand, is preparing its own model in collaboration with the creators of Cursor. OpenAI is set to publish full performance and pricing details for GPT-5.6 on the OpenAI API on launch day.

Sources: CNBC (cnbc.com), Nextgov/FCW (nextgov.com), Euronews (euronews.com), Quartz (qz.com)