Overview
- OpenAI announced GPT-5.6 on June 27, releasing a three-tier family called Sol, Terra and Luna plus a Sol Ultra variant that can add subagents to boost performance.
- The company said access will initially be limited to a small set of U.S. partners at the U.S. government's request, with broader availability planned in the coming weeks.
- OpenAI reported benchmark gains against competitor models on coding and biology tests and said Sol and Sol Ultra outscored a rival on Terminal-Bench 2.1 in its results.
- The company claimed large token-efficiency improvements for cybersecurity tasks, saying GPT-5.6 matches prior exploit-detection performance while using about one-third the tokens, and announced pricing meant to undercut rivals.
- OpenAI described new safety measures including real-time output classifiers, built-in jailbreak resistance and explicit refusals of forbidden cyber-assistance, and regulators' recent controls on Anthropic signal sustained government scrutiny that could affect who gets access and how firms build with the models.