OpenAI previews its next-generation AI models, led by GPT-4.6 Sol. This flagship model boasts significantly stronger capabilities in coding, science, and cybersecurity. Paired with its most advanced safety stack to date, Sol aims to bolster defenses against higher-risk activities and repeated misuse. The company is initiating a limited preview with trusted partners before broader availability.

Alongside Sol, OpenAI introduces Terra, offering competitive performance to GPT-4.5 at half the cost, and Luna, a fast, affordable option. All models feature enhanced safety measures, with specific attention paid to cybersecurity tasks. Sol's new 'max reasoning effort' allows deeper analysis, while an 'ultra mode' uses subagents for complex workflows. The model sets new benchmarks in coding and biology analysis, outperforming previous versions.

For cybersecurity, GPT-4.6 Sol shows improved performance-efficiency on long-horizon tasks, outperforming GPT-4.5 on benchmarks like Exploit Bench² with fewer tokens. While Sol can identify bugs and exploitation primitives, OpenAI states it does not autonomously produce functional full-chain exploits under tested conditions. This phased release and robust safety system are designed to benefit defenders first.

The preview period will involve close coordination with partners to refine safeguards and ensure legitimate work is not unduly hindered. OpenAI is also developing enterprise-specific safety solutions. This cautious approach prioritizes responsible deployment of advanced AI capabilities.