OpenAI previews GPT-5.6 Sol model with stronger safeguards

OpenAI has begun a limited preview of GPT-5.6 Sol, a new flagship model in its new GPT-5.6 family, which also includes Terra and Luna. The company said all three models are expected to become generally available in the coming weeks.

The company said the preview is initially limited to a small group of trusted partners. OpenAI said it shared its release plans and model capabilities with the US government before launch and is initially limiting access at the government’s request.

The company said it does not consider government pre-release access an appropriate long-term default. Instead, it described the limited preview as a temporary measure while working with the US administration on a repeatable release framework linked to a cybersecurity Executive Order.

OpenAI described GPT-5.6 Sol as its most capable model to date, highlighting improvements in agentic coding, biology and cybersecurity while saying a broader set of evaluation results will be published when the model becomes generally available.

For coding, OpenAI said GPT-5.6 Sol set a new state of the art on Terminal-Bench 2.1, which tests command-line workflows involving planning, iteration and tool coordination.

The company also reported improvements in biology workflows. On GeneBench v1, which evaluates long-horizon genomics and quantitative biology tasks, OpenAI said the model performed better than GPT-5.5 while using fewer tokens.

Cybersecurity is a major focus of the preview. OpenAI said GPT-5.6 Sol is its most capable model yet for cybersecurity tasks, including vulnerability research and exploitation-related workflows.

OpenAI said the model performs better at identifying and helping remediate vulnerabilities than at carrying out end-to-end offensive cyber operations. According to the company, GPT-5.6 Sol did not exceed the Cyber Critical threshold under its Preparedness Framework.

OpenAI said the GPT-5.6 release includes its most robust safeguards to date, with configurations tailored to each model’s capabilities. The company said these safeguards are intended to constrain prohibited offensive use while preserving access for legitimate work such as code review, vulnerability research, patch development, debugging, security education and defensive testing.

Safeguards include model-level protections, real-time generation checks, account-level monitoring, differentiated access controls, enforcement mechanisms and ongoing testing. OpenAI said some higher-risk requests may be delayed or blocked during the preview period.

The company said it devoted more than 700,000 A100-equivalent GPU hours to automated red-teaming, complemented by third-party expert testing, to evaluate the model’s resilience against jailbreak attempts.

During the preview, GPT-5.6 models will initially be available through the API and Codex to selected trusted partners and organisations. OpenAI said broader access for ChatGPT, Codex and API users is planned soon.

During the preview, GPT-5.6 models will be available through the API and Codex to selected partners. OpenAI said broader access across ChatGPT, Codex and the API is planned soon. It also announced pricing for the model family and said GPT-5.6 Sol will launch on Cerebras in July, initially for a limited group of customers.

Why does it matter?

GPT-5.6 Sol illustrates how frontier AI releases are becoming increasingly governed by phased deployment, targeted access and extensive safety testing rather than immediate public availability. OpenAI’s emphasis on cybersecurity evaluations, automated red-teaming and layered safeguards reflects growing efforts to manage the risks associated with increasingly capable foundation models.

The rollout also highlights the evolving relationship between AI companies and governments. By combining limited pre-release access, enterprise deployment and structured safety frameworks, OpenAI is helping shape emerging norms for how advanced AI systems are evaluated, governed and introduced into real-world use.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!