OpenAI has opened a limited initial preview of its GPT‑5.6 series, featuring three variants: Sol, the flagship model; Terra, a cost‑effective workhorse; and Luna, a fast, low‑price option. Sol arrives with the most extensive safety stack the firm has built, targeting high‑risk queries, cyber‑related requests, and repeated misuse. Access begins with a small group of trusted partners under U.S. government oversight.

Early tests show Sol outpacing GPT‑5.5 on several frontiers. On Terminal‑Bench 2.1, it set a new state‑of‑the‑art for command‑line planning and tool coordination. Gene Bench v1 revealed stronger genomics analysis while using fewer tokens, and Exploit Bench² placed Sol on par with Mythos Preview despite generating roughly one‑third the output. All three models also posted significant gains on UC‑Berkeley’s Exploit Gym.

The preview doubles as a stress test for the layered safeguards OpenAI built: model‑level refusals, real‑time misuse classifiers, account‑wide monitoring, and differentiated access tiers. When a high‑risk request triggers a pause, a larger reasoning model reviews the context before any output reaches the user. OpenAI plans to roll the models out broadly in the coming weeks, pending feedback from this trial.