OpenAI has updated GPT-6 Luna in ChatGPT, saying the October 29 version gives more concise responses and uses less overly agreeable language. Its accompanying system-card addendum reports improvements on some safety evaluations alongside higher factual-error rates than the October 7 version across six reported factuality measures.

The addendum is OpenAI’s own evaluation of the update. The figures describe its test sets and are not a guarantee of how the model will behave in every conversation.

On this page

What changed in the model

OpenAI calls the revision GPT-6 Luna (October 29). It says the update uses the same types of data and training as the October 7 version and the same safeguards stack. The company classifies both versions as High in cybersecurity and biological and chemical capability under its Preparedness Framework.

OpenAI says the update is intended to make responses more concise and reduce overly agreeable language. The system card also reports results across safety, prompt-injection resistance, health and factuality evaluations. The results are mixed. Some scores fell.

The safety results

On OpenAI’s production benchmark of challenging prompts, the reported gore score rose from 0.812 for the October 7 version to 0.858 for the October 29 version, with higher scores better. The company also marks improvements in four under-18 evaluation categories: sexual content, eating disorders, emotional reliance and gore.

For indirect prompt injection, the reported GPT-RED robustness score fell from 95.80% to 94.50%. OpenAI describes this table as an equal average across available reasoning levels. A separate direct instruction-hierarchy evaluation reached 99.91% robustness, according to the addendum. These are different evaluations and should not be read as one overall safety score.

The company says an attacker run for its multiturn jailbreak evaluation was cut short by infrastructure issues. It reports higher observed defender success for the updated model at every tested attacker budget, with that limitation attached.

Factuality and health evaluations

The addendum says factual-error rates increased against the October 7 version across all six reported factuality measures. OpenAI also says the new version remained below GPT-5.6 Luna’s error rates on each of those measures. That comparison is limited to the evaluations and versions described in the card.

On the reported length-adjusted HealthBench overall measure, the score moved from 50.4 to 48.2. The addendum reports a separate improvement on a dynamic self-harm evaluation and a minor regression on its mental-health evaluation. Different tests measure different tasks; none establishes that the model is suitable as a substitute for professional care.

What the release does not establish

The card describes a model update and internal test results. That is a limited view. It does not turn those figures into a prediction for every user, prompt or deployment. OpenAI notes that comparison values for previously launched models can reflect later versions and may differ from values published at their original launch.

GPT-6 Luna’s release dates and other OpenAI models are listed in AI model updates, and the week’s other model news is in AI news this week, October 24–30. For the broader distinction between training and serving models, read AI inference vs training compute.

Source