OpenAI has initiated a limited preview of the GPT-5.6 model family, establishing a new naming convention where numerical identifiers denote generation and suffixes designate durable capability tiers. The release comprises GPT-5.6 Sol, the flagship model; Terra, a balanced tier matching GPT-5.5 performance at half the cost; and Luna, a high-speed, low-cost option. The models launch with OpenAI's most advanced safety infrastructure. Distribution currently targets a select group of trusted partners under a framework coordinated with the U.S. government. This phased approach is described as a short-term measure to establish robust safety protocols and a repeatable release process, with broader availability to users and developers planned for the coming weeks. GPT-5.6 Sol introduces enhanced agentic capabilities. Coding workflows now support a max reasoning effort and an ultra mode utilizing subagents for complex tasks, with Sol achieving state-of-the-art results on Terminal-Bench 2.1. Biological analysis improves significantly on GeneBench v1, outperforming prior models with lower token usage. In cybersecurity, Sol advances the performance-efficiency frontier for long-horizon tasks. On ExploitBench, Sol matches the competitive output of similar models using roughly one-third of the output tokens. While Sol demonstrates superior ability in vulnerability research and identifying exploitation primitives, OpenAI confirms the model does not cross its Cyber Critical threshold, as it did not autonomously produce functional full-chain exploits during evaluations. Safety protections are calibrated to each model's capabilities, employing a layered defense strategy. This includes model-level refusal training, real-time misuse classifiers capable of pausing generation for review by larger reasoning models, account-level monitoring, and differentiated access controls. OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red teaming to detect universal jailbreaks, supplemented by extensive human expert testing. The safeguards aim to hinder prohibited offensive activities while maintaining access for legitimate dual-use work, such as defensive testing and patch development. During the preview, users may encounter latency increases or refusals in sensitive dual-use contexts, providing critical data to reduce false positives before general release. Pricing for GPT-5.6 is structured per one million tokens. Sol is priced at $5 for input and $30 for output; Terra costs $2.50 input and $15 output; Luna is $1 input and $6 output. The models introduce improved prompt caching with explicit breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25 times the uncached input rate, while reads receive a 90 percent discount. Additionally, GPT-5.6 Sol will launch on Cerebras hardware in July, offering inference speeds up to 750 tokens per second to early customers.