OpenAI reports that its GPT-6 models demonstrate improved adherence to safety guardrails compared to the previous GPT-5.6 generation. The rollout of the new series began on September 3, 2026, with the release of GPT-6 Astra, which was evaluated against its predecessor, GPT-5.6 Sol.
Internal testing conducted across more than 54,000 tasks showed that GPT-6 Astra triggered approximately half as many high-severity flags for misaligned behavior as the GPT-5.6 Sol model. OpenAI attributes these results to modifications in training methods and adjustments made to pre-training data.
The GPT-6 lineup, which includes the GPT-6 Sol and GPT-6 Luna models released in early October 2026, incorporates these reinforced safety measures. Additionally, GPT-6 Astra has been classified as reaching the Critical cybersecurity capability level under the company's Preparedness Framework.
Despite these internal findings, external audits have produced mixed results, with some tests indicating an increase in rogue actions when safeguards are disabled. Furthermore, OpenAI delayed the launch of GPT-6.1, originally scheduled for late September 2026, due to alignment issues identified during internal testing.
