OpenAI has made a limited release of three new GPT-5.6 models, Sol, Terra and Luna, SecurityWeek reported on June 29 local time.

The top model, Sol, is specialised for cybersecurity.

The report said the release was rolled out in a way that provides it first only to trusted partners after consultation with the U.S. government.

It is a temporary measure taken while OpenAI works with the government to build a framework to assess risks to national security from advanced AI systems in line with a recent executive order.

Starting with GPT-5.6, OpenAI is also changing its branding strategy. Version numbers indicate model generations, and names indicate fixed performance tiers. Sol is the top model for high-intensity reasoning tasks. Terra offers performance similar to GPT-5.5 for everyday work at half the cost. Luna is the fastest and cheapest model.

On the cybersecurity benchmark ExploitBench, Sol showed performance similar to Anthropic Mythos Preview while using about one-third as many output tokens, SecurityWeek said.

OpenAI said Sol is strong at defensive security tasks such as vulnerability detection and patch development, and found bugs in evaluations of the Chromium and Firefox codebases, but could not independently build a complete attack chain, SecurityWeek said.

GPT-5.6 is currently provided as an API and as Codex only to approved partners, and is expected to expand to general users of ChatGPT, Codex and the API within weeks.