GPT-5.6 enters preview under escort.

OpenAI launches GPT-5.6 with Sol, Terra, and Luna models in a secure US government-reviewed preview on API and Codex, starting at $1 per million tokens.

The GPT-5.6 family is launching with three models: Sol for the high-end, Terra for a cost-performance compromise, and Luna for faster and cheaper uses. Access begins with a limited preview in the API and Codex, reserved for a group of trusted partners, before a planned expansion to ChatGPT, Codex, and the API. OpenAI states it presented its plans and the models' capabilities to the US government beforehand, while specifying that this type of process should not become the norm.

In terms of capabilities, Sol introduces a `max` reasoning effort and an `ultra` mode, based on the use of subagents for complex tasks. OpenAI highlights progress in coding, biology, and cybersecurity, with results superior to GPT-5.5 on Terminal-Bench 2.1, GeneBench v1, ExploitBench, and ExploitGym according to its own evaluations. In cybersecurity, the company emphasizes one point: Sol would be more useful for finding and fixing vulnerabilities than for single-handedly conducting full attacks against hardened targets.

The preview is therefore accompanied by a reinforced security layer: trained refusals within the model, real-time checks during generation, analysis of account-level signals, differentiated access, and automated red-teaming. OpenAI states it dedicated over 700,000 A100e GPU hours to searching for universal jailbreaks. For API pricing, Sol is announced at $5 for input and $30 for output per million tokens, Terra at $2.50 and $15, and Luna at $1 and $6. Access to Sol on Cerebras is also planned, with an announced speed of up to 750 tokens per second for certain clients.