Less than 24 hours after news emerged that OpenAI would stagger the release of its next model at the request of the Trump administration, the tech company has unveiled its new GPT-5.6 model suite. Announced on Friday, this limited preview features three distinct models: Sol, Terra, and Luna. Sol represents the flagship model with advanced capabilities, while Terra is designed for medium-tier applications and high-volume workloads. Luna, on the other hand, is characterized as a “fast and affordable” option for everyday use.
OpenAI has highlighted GPT-5.6’s strengths in various domains, including coding, cybersecurity, and biology, and emphasizes its ability to maintain focus on long-term, agentic AI tasks. The pricing structure is notably competitive, with Sol available for $5 input and $30 output per million tokens—almost half the cost of Anthropic’s Claude Fable 5 model, which is priced at $10 input and $50 output. Terra is offered at half the price of Sol, while Luna costs even less than Terra.
In addition to the models, OpenAI has introduced two new operation modes for Sol: a “max” mode for enhanced reasoning capabilities and an “ultra” mode that allows for leveraging sub-agents. This development alludes to the company’s attempts to incorporate innovations from Peter Steinberger, the creator of OpenClaw, who has notably contributed to OpenAI’s efforts.
Given recent security concerns in Washington, D.C., OpenAI devoted much of its blog post announcement to the safety measures and potential misuse of the technology. The company appears to have taken particular note of the difficulties faced by Anthropic with their models and stated that “GPT-5.6 is trained to refuse prohibited cyber assistance,” aiming to prevent users from disguising malicious intents or jailbreaking the model. It further asserted that Sol excels in identifying and addressing vulnerabilities rather than executing end-to-end attacks and assured users that it does not breach the cyber-critical threshold under OpenAI’s revised preparedness framework.
The company claims that Sol is equipped with its “most robust safety stack to date,” implementing strengthened protections for higher-risk activities, sensitive cyber requests, and instances of repeated misuse. Notably, OpenAI mentioned it dedicated “approximately 700,000 A100e GPU hours” to automated red-teaming, and engaged with third-party testers who will continue their assessments over the next two weeks.
During this preview period, which is closely monitored by the Trump administration, OpenAI indicated that safeguards may sometimes disrupt legitimate work, especially in dual-use scenarios where defensive and offensive activities may appear similar. In light of this, the administration is set to approve customers on a case-by-case basis during the preview.
OpenAI expressed confidence that the model suite would be generally available in the coming weeks, advocating for “broad access.” The company did confirm its cooperation with the US government prior to the launch but expressed hope that such measures would not become standard. “We don’t believe this kind of government access process should become the long-term default,” the company stated. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” OpenAI emphasized that this temporary arrangement was aimed at facilitating a wider availability of the models in the near future while collaborating with the administration to establish a framework and a repeatable process for future releases.
