GPT-5.6 Sol is OpenAI's flagship large language model, released as the top tier of the three-model GPT-5.6 family alongside Terra (a balanced, lower-cost option) and Luna (a fast, cost-efficient variant). It is positioned as OpenAI's most capable model for complex reasoning, coding, scientific research, cybersecurity, and long-horizon agentic work. The model follows a limited preview period in late June 2026, during which access was coordinated with the U.S. government due to the model's cybersecurity and biological-risk capability profile, before reaching general availability.
Architecture and Capabilities
GPT-5.6 Sol introduces a new "max" reasoning effort setting that allows the model additional time to reason before responding, along with an "ultra" mode that coordinates multiple parallel subagents to distribute complex tasks across concurrent workers. The model supports a context window of roughly 1.05 million tokens with a maximum output of 128,000 tokens, and accepts both text and image inputs. It also supports tool use, code execution, structured outputs, and prompt caching, with programmatic tool calling available through the Responses API.
Performance
OpenAI reports that Sol, using max reasoning, achieves state-of-the-art results on the Artificial Analysis Coding Agent Index and on Terminal-Bench 2.1, a benchmark for command-line coding workflows. It also achieves new results on BrowseComp, a benchmark for agentic browsing tasks, and on OSWorld 2.0, a computer-use benchmark. On the Agents' Last Exam evaluation, which spans long-running professional workflows across 55 fields, OpenAI reports a score of 53.6.
Safety
Under OpenAI's Preparedness Framework, GPT-5.6 Sol is classified as High capability in both cybersecurity and biological/chemical risk domains, though it does not reach the framework's Critical threshold. In cybersecurity evaluations, the model was able to identify vulnerabilities and exploitation primitives but did not autonomously complete full end-to-end attacks against hardened targets under tested conditions. OpenAI paired the release with additional safeguards, including activation classifiers for sensitive domains and programs that reserve certain advanced capabilities for vetted defensive users.
Exact parameter counts and training architecture details have not been publicly disclosed.