OpenAI tunes GPT-6 Luna for shorter, less sycophantic answers

OpenAI pushed an update to GPT-6 Luna focused on tone: answers are more concise, and the model curbs the overly agreeable language users have long called sycophancy. The release came with a deployment safety addendum dated October 28.

That addendum classifies Luna as High for cybersecurity and biological capabilities under OpenAI's Preparedness Framework, which triggers extra safeguards before deployment. Publishing the rating alongside a personality tweak shows how even small model updates now go through formal safety review.

The update follows the broad availability of GPT-6.1 Terra for visual reasoning and computer use across ChatGPT Work and Codex. Sycophancy has become a competitive issue: Anthropic this week published research showing analysts matched Claude's nudged preference 86% of the time, highlighting how agreeable assistants can steer decisions. A less flattering model is a direct response to that concern.

The developer reaction is mixed. Many welcome the end of conversational drift and flattery, but some frustration in the community is aimed at OpenAI's broader product moves, including wider ad formats and a Reddit thread on merging chat and work usage limits. Ethan Mollick noted OpenAI often retires products within months, a reminder that today's model behavior can change quickly.