“On the ExploitBench cybersecurity benchmark, OpenAI's GPT-5.6 SOL model achieves performance comparable to Anthropic's Mythos while using approximately one-third of the tokens.”