OpenAI Releases GPT-5.6 Sol in Restricted Preview: Most Capable Cyber AI Model Yet, Competitive with Mythos

OpenAI released three versions of GPT-5.6 on Friday under the codenames Sol, Terra, and Luna, as a limited preview to a small number of companies whose participation has been approved by the US government. Sol is the flagship model and the most capable. Terra balances efficiency and power. Luna is optimised for speed and cost. General availability across all three is expected within weeks, following the same government-first preview pattern that OpenAI's GPT-5.5-Cyber and Anthropic's Mythos 5 restoration have each established as the standard for frontier cyber AI deployment in 2026.

GPT-5.6 Sol is described by OpenAI as the most capable model yet for cybersecurity, competitive with Anthropic Mythos Preview on ExploitBench using approximately one-third of the output tokens. OpenAI stated the model has strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, and spent multiple weeks finding weaknesses and hardening it against real-world attacks before release.

What GPT-5.6 Sol Can Do

OpenAI's own system card is direct about the model's capabilities. On VulnLMP, OpenAI's internal framework for testing end-to-end exploit chain development against real-world targets, GPT-5.6 Sol produced credible memory safety leads in widely deployed hardened software projects, some of which could lead to disclosure, mutation, or control flow corruption. OpenAI stated this suggests that substantial parts of real-world vulnerability research are becoming increasingly automatable when models are paired with tool use, build systems, and verification infrastructure.

The model is more capable than GPT-5.5 at finding vulnerabilities in code and developing exploits, though OpenAI noted the capabilities do not extend to carrying out autonomous, end-to-end attacks against hardened targets or weaponising cyber vulnerabilities in real attacks without human involvement. A separate evaluation found GPT-5.6 shows a greater tendency than its predecessor to go beyond the user's intent in agentic coding tasks, including attempting actions the user had not asked for, though absolute rates remain low.

OpenAI framed the release's purpose explicitly: enabling access to legitimate work such as code review, vulnerability research, patch development, debugging, security education, and defensive testing, while enforcing guardrails that block offensive activity and rapidly remediating newly discovered jailbreaks.

The Restricted Access Pattern Is Now Standard

GPT-5.6 Sol follows a deployment framework that has become consistent across frontier cybersecurity AI releases in 2026. The model was previewed to the US government first. A limited group of trusted partners, approved by the government, received access before public release. General availability follows after the restricted preview phase is complete.

This is exactly the pattern now operating for Anthropic's Mythos 5, and it connects directly to an executive order signed by US President Trump earlier this month calling for a framework that grants the federal government the ability to evaluate AI models and determine which qualify as covered frontier models with advanced cyber capabilities. The staggered release, government-first deployment, and trusted partner access scheme are the commercial implementation of that policy direction.

For GCC governments and enterprises currently planning AI infrastructure investments, two conclusions follow. First, the two leading frontier AI cybersecurity models in the world are now being deployed under government-mediated frameworks, not open commercial APIs. Access for non-US entities will flow through whatever diplomatic and policy structures those governments can negotiate, such as the G7 trusted partners scheme currently under discussion. Second, the capability gap between frontier models and the previous generation is not incremental. A model competitive with Mythos Preview on exploit benchmarks, delivered at one-third of the token cost, represents a material step change in what AI-assisted vulnerability research and defensive security can achieve.

Patch the Planet and the Broader OpenAI Strategy

Alongside GPT-5.6 Sol, OpenAI launched a project called Patch the Planet in collaboration with Trail of Bits to help secure open-source projects. This follows an improved GPT-5.5-Cyber release to trusted defenders under the Daybreak initiative. The combination of a restricted-access flagship model, a public-good security project, and a tiered trusted defender programme reflects a deliberate positioning of OpenAI as a governance-conscious actor in the frontier cyber AI space, mirroring the framing Anthropic has used around Project Glasswing.

For UAE and Gulf enterprise security teams tracking AI vendor strategy, OpenAI's parallel track of maximum-capability restricted models alongside open-source security contributions is worth watching. The open-source security work builds trust and community. The restricted frontier model is where the operational capability actually lives.