OpenAI launches GPT-5.6 series with strong cybersecurity capabilities, but access remains limited for now

On 26 June 2026, OpenAI launched three new models under the collective name GPT-5.6: Sol, Terra and Luna. Sol is the most capable of the three and performs well on coding, biology workflows and identifying software vulnerabilities. Terra is intended for everyday use, while Luna is the faster and more affordable option.

The models are currently available only to a small circle of trusted partners, all based in the United States. That restriction comes at the request of the Trump administration, which wants to review the models' advanced capabilities before broader access is granted. OpenAI presented GPT-5.6's plans and performance to the US government ahead of the launch. The names of the participating partners have been shared with the authorities.

The company expects to make the models generally available in the coming weeks and hopes to add international partners to the preview as early as next week.

What Sol can do in cybersecurity and biology

GPT-5.6 Sol scores comparably to Anthropic's Mythos Preview on the ExploitBench² benchmark, a model also known for its performance in identifying vulnerabilities and generating exploits. Sol reaches that level using approximately one-third of the output tokens that Mythos requires.

On Terminal-Bench 2.1, a measure of command-line workflows requiring planning, iteration and tool coordination, Sol achieves new top results relative to previously published scores. The model also features a so-called max reasoning effort setting that enables deeper reasoning, and an ultra mode that deploys sub-agents to accelerate complex tasks.

In the area of biology, Sol scores higher than GPT-5.5 on GeneBench v1. On the SecureBio evaluations, which measure how well a model handles biologically sensitive knowledge, Sol achieved scores of 53.5% on the Virology Capabilities Test, 60.0% on Molecular Biology, 68.4% on Human Pathogen Capabilities and 68.3% on a further biosecurity category. Those scores are on average approximately nine percentage points higher than GPT-5.5. This partly explains why the US government wants to closely monitor the model's rollout.

How the access restriction came about

The situation with GPT-5.6 resembles what Anthropic experienced earlier with its Mythos 5 models. Those were temporarily blocked by the US government for international users on national security grounds. On the same day OpenAI launched GPT-5.6, the government lifted the block on Anthropic's Mythos 5: the model is now accessible to more than a hundred US institutions, including large corporations and government agencies.

OpenAI is working with the government on a framework around a so-called cyber Executive Order and on a repeatable process for future model releases. CEO Sam Altman has communicated internally about the phased rollout. Chief Scientist Jakub Pachocki reportedly briefed employees on the scale of the step that GPT-5.6 represents relative to earlier models. George Osborne, OpenAI's head of countries, had previously announced that the company would give the US government the opportunity to review new models before they are broadly deployed.

OpenAI has publicly stated that it does not view this approval process as a permanent model. Withholding broadly available access to advanced tools from users and partners who need them is not consistent with the company's own objectives, it said.

What this means for European and Dutch users and businesses

For now, European users and businesses, including Dutch ones, remain outside the preview. OpenAI is aiming to add international partners in the week following the launch, but no concrete commitments have been made regarding timelines or which partners will be included.

For founders, investors and policymakers in Europe, the developments around GPT-5.6 are a concrete example of how national security considerations can affect the availability of AI tools, even for parties outside the US. The pattern of the US government having a say in who gains access to advanced models has now become visible twice in quick succession, first with Mythos 5 and now with GPT-5.6. European policymakers working on their own AI governance frameworks, including under the EU AI Act, may factor this mechanism into discussions around access, sovereignty and dependence on US providers.