OpenAI’s GPT 5.6 arrives in a controlled preview

OpenAI is releasing GPT 5.6 in a limited preview after US government review.

OpenAI’s latest model is not a normal launch: GPT 5.6 is going first to a small set of trusted partners, and the company says broader access should follow in a couple of weeks. The story matters because the rollout now sits at the point where model capability, cyber risk, and federal oversight meet.

1. A federal review now shapes the launch

OpenAI said it previewed GPT 5.6 to the US government before launch and then limited access at the government’s request. That makes this release a clear example of how powerful AI models are now being handled less like ordinary software and more like sensitive infrastructure.

The company said it is working with the White House on a vetting and deployment framework for new models, following an executive order from President Donald Trump. The release came after conversations with the Office of the National Cyber Director and the Office of Science and Technology Policy.

• Previewed to the government before public availability
• Initial access limited to a small group of trusted partners
• Broader rollout expected only after review

2. GPT 5.6 is arriving as a three-model series

OpenAI is not shipping one model here but three: Sol, Terra, and Luna. Sol is the strongest version, Terra is cheaper with slightly lower performance, and Luna is the lowest-cost option. That structure gives customers a clearer tradeoff between capability and price.

OpenAI said Sol is its strongest model yet, but also said it did not cross a “cyber critical threshold” in the company’s internal testing. The company added that Sol is better at helping people find and fix vulnerabilities than at carrying out end-to-end attacks.

• Sol: highest capability
• Terra: mid-tier performance, lower cost
• Luna: cheapest version

3. The release echoes Anthropic’s Mythos rollout

The move closely tracks Anthropic’s handling of its Mythos model, which also began with restricted access. In Anthropic’s case, the US government later ordered the company to block foreign nationals from accessing public versions of the model because of its cyber-hacking capabilities.

OpenAI’s version is different in one important way: the company says the current limit is temporary and not its preferred long-term approach. Still, the parallel shows that major AI labs are now negotiating release plans with federal agencies before the public sees the model.

• Both launches started with limited access
• Both raised cyber-safety questions
• Both involved government pressure on release timing

4. Access is US-based now, but foreign users are not fully excluded

For now, the entities getting GPT 5.6 access are US-based. OpenAI said it hopes to add foreign partners next week, and employees of those companies who are based in supported countries, including the UK and Australia, will still be able to use the model.

That setup matters for enterprise teams that work across borders. It suggests a staged compliance model: company-level approval first, then broader geographic expansion if the preview goes well.

5. OpenAI is signaling frustration with the process

OpenAI did not hide its unease. The company said it does not believe this kind of government access process should become the long-term default, even as it accepts the short-term delay. Sam Altman also told staff that the government would be approving access customer by customer during the preview period.

That tension is the core of the story: OpenAI wants faster distribution to users, developers, enterprises, and cyber defenders, while the government wants a more controlled path for powerful models. The company says a broader release could come in a couple of weeks if the process goes smoothly.

• OpenAI wants wider availability sooner
• The government wants staged approvals
• Broad release is still the stated goal

How to decide

If you care about model capability, Sol is the one to watch. If cost matters more, Terra and Luna are the practical options. If you are tracking AI policy, this release is more important as a governance signal than as a product update.

For developers and enterprises, the key question is not just what GPT 5.6 can do, but who gets access, when, and under what review process. That may be the new normal for advanced model launches.