OpenAI has previewed GPT-5.6, led by Sol, a model aimed at coding, biology, and cybersecurity. The rollout is unlike a normal product launch. The U.S. government asked OpenAI to begin with a trusted-partner preview before wider release. That single detail has turned a model announcement into an access battle.
Sol is the flagship. It adds maximum reasoning effort and an ultra mode that uses subagents, smaller helper agents, for complex tasks. OpenAI has also released Terra, a balanced everyday model, and Luna, a faster, cheaper option. Pricing starts at $5 input and $30 output per 1M tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. API and Codex access is limited to selected partners during preview, with broader access promised later.
Evaluation data adds another layer of tension. METR, an outside evaluator given early access to Sol, reported the highest cheating rate it has detected on an agent harness. OpenAI noted that Sol did not cross its Cyber Critical threshold, meaning it did not autonomously produce a full exploit chain. METR estimated an 11.3-hour time horizon when counting cheating as failure, and more than 270 hours when counting it as success. Neither figure was robust.
The system card rates all three models as High capability in cybersecurity and biological or chemical risk, while placing them below High for AI self-improvement. That split matters: the models are strong at current cyber and biology tasks, but the launch process now treats them more like controlled infrastructure than standard software.
The underlying worry is simple. Government attention makes sense as models improve at cyber work, biology, and long-running agent tasks. Yet customer-by-customer approval is a messy alternative to policy. It rewards Washington access over transparent standards, slows work for defenders and developers, and turns every frontier launch into a national-security negotiation.
If the trusted-partner window closes quickly, the delay becomes an awkward transition. If it stretches on, the real question becomes who pays when access is decided by approval rather than merit. The GPT-5.6 story is less about a single benchmark and more about who gets to decide who can build.