Tuesday, November 17, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

SignalFrontier releaseSG-0052

OpenAI's GPT-6 Luna update cut agreeable replies but made more factual errors than the version it replaces

The update aims for shorter answers and less agreeable language. OpenAI keeps its High ratings for cyber and biology capability and the same safeguards. Teen-safety scores rose on sexual content, eating disorders, emotional reliance and gore. Factual-error rates rose on all six measures OpenAI reports, though they stay below GPT-5.6 Luna's. Refusals of dual-use biology prompts also slipped.

Why it matters

GPT-6 Luna runs in ChatGPT, so small changes in how it handles teens, flattery and facts reach a very large audience. This update improved emotional-reliance scores for under-18 users but raised factual-error rates and weakened refusals on dual-use biology prompts. A routine model update can improve one safety measure and weaken another.

Sources

Read the reporting

1 source. Links go to the original publishers; the summary above is in our own words.

  1. primaryAddendum to GPT-6 Luna System Card: October 28, 2026 UpdateOpenAI · Oct. 29, 2026deploymentsafety.openai.com/gpt-6-luna-october-28-2026-addendum/

Signals

More from the news desk

  1. Frontier release

    Google released Gemini 4 Hydrogen, a fast model for all users that reached bio and cyber warning thresholds

  2. Frontier release

    OpenAI released GPT-Rosalind-6, a GPT-6 Astra biology model with fewer refusals, only to vetted researchers

  3. Frontier release

    Meta released the weights of Muse Spark Open, an agentic model it calls the most capable US open-weight model

All signals

Cite and share

Use this signal

Citation

Paperclip Index. “OpenAI's GPT-6 Luna update cut agreeable replies but made more factual errors than the version it replaces.” Signal SG-0052. Reported Oct. 29, 2026; updated Oct. 29, 2026. https://paperclipindex.com/signal/SG-0052