OpenAI Previews GPT-5.6 Sol

OpenAI has begun a limited preview of its next-generation large language models, including GPT-5.6 Sol, its new flagship model. This release also introduces Terra, a balanced model, and Luna, a fast and affordable option.

Terra aims to match GPT-5.5's performance at half the cost, while Luna offers strong capabilities at OpenAI's lowest price point.

Enhanced Capabilities and Safety

GPT-5.6 Sol represents OpenAI's most capable model to date, showcasing significant improvements in agentic capabilities across coding, biology, and cybersecurity workflows. It introduces a new max reasoning effort and an 'ultra mode' that leverages subagents for complex tasks.

The model sets new benchmarks in areas like Terminal-Bench 2.1 for command-line workflows and GeneBench v1 for genomics. In cybersecurity, GPT-5.6 Sol improves performance-efficiency for long-horizon tasks such as vulnerability research and exploitation, demonstrating competitive results on ExploitBench² and ExploitGym³.

Alongside these advances, OpenAI emphasizes the most robust safety stack to date. Safeguards are layered and tailored to each model's capabilities, designed to withstand real-world adversarial pressure while supporting legitimate defensive work.

GPT-5.6 Sol is assessed as beneficial for finding and fixing vulnerabilities rather than executing end-to-end attacks. It does not cross the 'Cyber Critical' threshold under OpenAI's Preparedness Framework.

Layered Safeguards and Phased Release

OpenAI employs a multi-faceted approach to safety, integrating protections directly into the model's training, implementing real-time generation checks, and utilizing account-level signals for misuse detection. This layered strategy includes differentiated access, monitoring, and continuous testing.

The company is starting with a limited preview for a select group of trusted partners, a step taken in coordination with the U.S. government. This phased release allows for further testing and refinement before broader availability.

OpenAI acknowledges that this government access process should not become a long-term standard, as it can restrict access to valuable tools. This short-term measure is intended to facilitate broader availability in the coming weeks, aligning with efforts to develop a cyber Executive Order framework.