OpenAI says GPT-6.1 Terra solves 50.3% of the Terminal-Bench 4.0 command-line coding tasks, up from GPT-5.6 Terra's 21.5%