The company has unveiled its flagship GPT-5.6 Sol and claims that in vulnerability discovery and exploitation, it performs at the level of Anthropic’s top model but uses three times fewer tokens.
On the ExploitBench benchmark, Sol is comparable to Claude Mythos Preview, using only about 1/3 as many output tokens.
The model is currently available only to a limited circle of trusted partners. OpenAI expects to open access in the coming weeks.
Sol is the company’s strongest model for cybersecurity: tackling vulnerability discovery and exploitation tasks.
The lineup also includes Terra (balancing performance and price) and Luna, the fastest and most affordable model.
Continue reading this article on source: openai.com