OpenAI updated GPT-6 Luna in ChatGPT on October 29 to give shorter answers and reduce overly flattering language. The change reaches Free and Go users and people chatting without signing in. For Plus and Pro subscribers, it affects only the fallback used after GPT-6 message limits are reached; their main model is unchanged. The accompanying system card identifies the revised model as GPT-6 Luna (October 29).
OpenAI reports stronger safety results in several categories for users under 18 and 99.91% robustness on its instruction-hierarchy evaluation, which tests resistance to attempts to override higher-priority instructions. The results are mixed elsewhere: its length-adjusted HealthBench score fell from 50.4 to 48.2 on a 0–100 scale. More concise answers therefore arrive alongside measurable trade-offs in the developer's own evaluations.
Source · deploymentsafety.openai.com ↗
