OpenAI has announced a limited preview of its new GPT-5.6 model lineup, introducing three tier-based systems named Sol, Terra, and Luna. Sol has been described by the AI giant as its most capable model for cybersecurity.
The initial launch follows a consultation process with the United States government, resulting in a restricted early rollout to a select group of trusted partners.
OpenAI indicated that this phased release is a temporary measure while the organization collaborates with the federal administration to develop a framework aligned with the recent Executive Order, which enables the government to assess the national security risks posed by advanced AI systems.
The GPT update marks a shift in OpenAI’s branding strategy, where version numbers designate the model generation, while the names represent permanent capability tiers. GPT-5.6 Sol is the premier flagship model, designed for high-intensity reasoning tasks. GPT-5.6 Terra targets everyday workloads, with OpenAI reporting performance competitive with GPT-5.5 at half the operating cost. GPT-5.6 Luna is positioned as a fast alternative representing the lowest pricing point in the lineup.
In cybersecurity tests on ExploitBench, Sol matched competing systems like Mythos Preview while using roughly a third of the output tokens.
The company emphasized that Sol demonstrates a stronger affinity for defensive security tasks, such as vulnerability identification and patch development, rather than end-to-end cyberattacks. During evaluations on the Chromium and Firefox codebases, the system successfully isolated bugs and basic exploitation primitives but did not independently construct a working full-chain exploit.
To manage potential dual-use risks where defensive research closely mirrors offensive behavior, the GPT-5.6 series relies on a multi-layered security architecture. In addition to standard training-level refusals, the infrastructure employs automated real-time classifiers for biology and cybersecurity inputs. If an anomaly is flagged, output generation pauses while a secondary reasoning model reviews the context of the conversation.
Some instances may also trigger account-level evaluations to help distinguish legitimate security research from malicious behavior.
OpenAI said it allocated over 700,000 A100-equivalent GPU hours toward automated red-teaming to harden the model against systemic vulnerabilities. This approach focused on discovering universal jailbreaks rather than single-prompt failures.
While the models are currently accessible only via the API and Codex to approved partners, OpenAI plans to expand general availability to ChatGPT, Codex, and broader API users in the coming weeks.
The company publicly pushed back against making government-mediated pre-clearance a long-term standard for AI deployments, arguing that prolonged restrictions delay essential defensive tools from reaching the broader cybersecurity community.
Related: Anthropic Launches Claude Fable 5: Mythos-Class AI With Cybersecurity Guardrails
Related: Claude Mythos Turns N-Days Into N-Hours With Rapid Exploit Creation
Related: New Enterprise-Ready MCP Specification Brings New Security Challenges
