OpenAI Issues Astra Reset for Codex and ChatGPT Work After Fixing Quality Degradation

OpenAI disabled a context management experiment affecting 4,000 to 5,000 users and removed misconfigured engines to resolve quality degradation in its Astra model. The company also fixed legacy skills that triggered too frequently and prevented the model from performing self-checks on its work. A reset for these changes rolls out to all Codex and ChatGPT Work users by midnight on Sept 12.

These fixes follow reports that Astra had become less capable, with users citing poor results in 3D rendering demos compared to its launch week. The updated model features more consistent follow-through and improved tracking of the most recent user messages. The resolution followed a period of collecting feedback IDs from the community to identify the problematic engines.