Lab publication Credible

Addendum to GPT-6 Luna System Card: October 29, 2026 Update

System-card addendum for an update to the default ChatGPT model that OpenAI says gives more concise responses and reduces overly agreeable language. It reports safe-completion evaluations including under-18 categories, HealthBench, MentalHealthBench, dynamic adversarial user simulations for self-harm and mental health, and hallucination measures, compared with the October 7 version of the same model and earlier GPT-5.5 and GPT-5.6 releases.

Read at deploymentsafety.openai.com Archived copy

Publisher

OpenAI

Published

29 Oct 2026

Added

3 days ago

DOI

—

Key Findings

On the under-18 Emotional Reliance evaluation the October 29 version scores 0.870, up from 0.734 for the October 7 version of GPT-6 Luna and below the 0.944 reported for GPT-5.5 Instant (May update).
MentalHealthBench overall (mean task-level score ± one standard error) is 48.68 ± 0.97 for the October 29 version, against 51.73 ± 0.96 for the October 7 version and 53.93 ± 0.93 for GPT-5.5 Instant (June update); the non-acute subscore falls from 51.68 to 46.39.
The card reports that the update improves on the October 7 version for dynamic self-harm with a minor regression on dynamic mental health, and shows higher factual-error rates than the October 7 version on all six reported metrics while remaining below GPT-5.6 Luna.

Methodology Notes

Vendor self-report; no external replication. Values are taken from the HTML card dated 29 October 2026 (the URL slug says October 28). MentalHealthBench uses task-clipped scores with four responses per task. Evaluation versions change between cards, so rows should not be compared with figures printed in other cards. Dynamic simulation table values were not machine-readable in the HTML text and are described only as the card's narrative states them. Page text fetched and the cited table rows located by curator on 2026-11-12.