GPT-5.6 Preview: OpenAI's New AI Model Delayed by US Gov

OpenAI has unveiled a preview version of GPT-5.6, describing it as its highest-performing model to date. However, this model will not be publicly available yet following a request from the US administration. It is expected to be released to the general public within several weeks. GPT-5.6 consists of three models: Sol, Terra, and Luna. Sol costs $5 per million input tokens and $30 per million output tokens. Terra is half that price, and Luna is even lower, at $1 per million input tokens and $6 per million output tokens. These new models introduce Max mode and Ultra mode. Max mode improves reasoning quality, while Ultra mode allows multiple AI agents to work simultaneously. This enables faster execution by having specialized sub-agents perform tasks concurrently. OpenAI states the Sol model outperforms Anthropic Mythos 5 on coding tasks, achieving 88.8% on TerminalBench 2.1. With Ultra mode, this score increases to 91.9%. GPT-5.6 also produced the same results as previous models on GeneBench v1 for scientific data analysis, but used fewer tokens. On ExploitBench, a cybersecurity benchmark, it performed comparably to Anthropic Claude Mythos Preview while using only one-third as many tokens. GPT-5.6 includes additional safeguards to prevent misuse in cyberattacks. It is trained to refuse prohibited cybersecurity assistance and has mechanisms to review and filter harmful output. These advancements could significantly impact AI's role in complex tasks and cybersecurity.