OpenAI has released an early preview of its latest artificial intelligence models in the GPT-5.6 series. The series includes three versions meant to fit different levels of need and cost. The main version, called Sol, is presented as the strongest of the group. It shows gains on tasks that require careful, extended thinking such as writing computer code, exploring scientific topics, and addressing protection against online threats. To support this, the model can apply more computing power to its thinking steps and can divide hard problems into smaller parts handled by helper systems that later combine their results.
A second version called Terra aims to deliver performance close to the prior main model while cutting the cost roughly in half for the same amount of work. The third version, Luna, is built for speed and low cost when processing very large numbers of requests. All three use an updated way of storing earlier inputs so that similar new questions can be answered faster and with less computing effort.
Safety measures and early availability
OpenAI reports stronger checks to reduce risks of harmful uses, including tests focused on digital attacks and other sensitive areas. The company states that it spent weeks searching for weaknesses before releasing the preview. Due to requests from United States government officials, the models are first limited to a small set of trusted partners. OpenAI plans to expand access to more users in the coming weeks. Pricing is set so that the most capable version costs more per unit of text processed while the faster versions cost less.
The announcement presents these changes as part of ongoing work to improve how well the systems handle real tasks that unfold over many steps. It also notes that full test results and safety details appear in a separate document. The overall approach keeps the focus on measured progress in capability while adding controls on who can use the models at the start. These steps reflect a pattern of steady updates rather than sudden large jumps in what the systems can do.