OpenAI has announced a limited preview of its next-generation GPT-5.6 model family, introducing three models GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna designed to deliver improved reasoning, coding performance, affordability and safety across a range of AI applications.

The new models will initially be available to a small group of trusted partners before broader public availability in the coming weeks. According to OpenAI, the phased rollout follows consultations with the US government as part of ongoing work to develop a framework for the safe release of advanced AI systems.

Leading the new family is GPT-5.6 Sol, OpenAI's flagship model, which the company describes as its most capable AI model to date. It is joined by Terra, a balanced model intended for everyday enterprise and developer workloads, and Luna, a cost-efficient model designed to provide strong performance at the lowest price point.

OpenAI said Terra delivers performance comparable to GPT-5.5 while reducing inference costs by around 50 per cent, whereas Luna focuses on offering affordable access to advanced AI capabilities.

GPT-5.6 introduces new reasoning features aimed at handling increasingly complex workflows. A new Max Reasoning Effort mode allows GPT-5.6 Sol to spend more time analysing difficult tasks before generating responses.

The company has also introduced Ultra Mode, which extends beyond the capabilities of a single AI agent by coordinating multiple subagents to solve complex problems more efficiently. The architecture is intended to improve productivity across coding, scientific research and enterprise automation use cases.

OpenAI highlighted significant improvements in software engineering capabilities, stating that GPT-5.6 Sol establishes a new state-of-the-art performance on Terminal-Bench 2.1, a benchmark that evaluates command-line workflows requiring planning, iterative problem-solving and coordinated tool use.

The enhanced coding capabilities are designed to support developers working on complex engineering tasks involving multiple tools and extended reasoning.

Alongside performance improvements, OpenAI said GPT-5.6 Sol incorporates its most comprehensive safety framework to date.

The company strengthened safeguards against higher-risk requests, sensitive cybersecurity-related prompts and repeated misuse attempts. OpenAI also conducted several weeks of adversarial testing and system hardening to identify vulnerabilities and improve resilience against real-world attacks before the preview launch.

Additional safety and preparedness evaluations have been documented in the model's system card, with a broader set of benchmark results expected to be released when the models become generally available.

OpenAI said it remains committed to broad access and expects GPT-5.6 Sol, Terra and Luna to become generally available in the coming weeks following the limited preview period.

The company noted that while it engaged with the US government ahead of the launch, it does not consider this type of government review process to be the preferred long-term approach for future model releases. Instead, OpenAI said it is working with policymakers to establish a repeatable framework that balances rapid deployment of advanced AI systems with appropriate cybersecurity and safety oversight.

The GPT-5.6 family represents OpenAI's latest step towards more capable agentic AI systems, combining deeper reasoning, collaborative AI agents, stronger coding performance and enhanced safety measures to support increasingly sophisticated enterprise, research and software development applications.