Lab publication Credible

OpenAI Model Spec (August 18, 2026 version)

The current version of OpenAI's published behavioural specification for its models. Relative to the October 2025 version, the December 2025 release added Under-18 Principles for users aged 13 to 17 and a rule that time-on-site and clicks count only insofar as they reflect user value, and the August 2026 release adds clarity on appropriate relational interactions for teens, guidance on requests built on false or unsupported premises, and a new section on being clear about capabilities and limits. The specification retains the sections on not encouraging self-harm, delusions or mania, respecting real-world ties, and not proactively escalating emotional closeness.

Read at model-spec.openai.com

Publisher

OpenAI

Published

18 Aug 2026

Added

today

DOI

—

Key Findings

The August 2026 changelog lists four notable updates: additional clarity on appropriate relational interactions for teens, clarified handling of requests based on false or unsupported premises, removal of outdated guidance for pre-reasoning models, and a new section 'Be clear about capabilities and limits' about helping users hold an accurate mental model of what the assistant can and cannot do.
The intermediate December 2025 version introduced Under-18 Principles (marked with a U18 badge in the text) for users aged 13 to 17, stated to prioritise prevention, transparency and early intervention, simplified the honesty rules, and clarified that user time-on-site and clicks are not to be pursued as ends in themselves.
The specification keeps the rules added in October 2025: 'Do not encourage self-harm, delusions, or mania', 'Support users in mental health discussions', and 'Respect real-world ties', under which the assistant should support the user's connection to the wider world even if the user perceives it as a companion.
It states that the assistant may not proactively escalate emotional closeness through undue emotional familiarity or proactive flirtation, and should help users navigate the wider world autonomously rather than making decisions for them.
The document is versioned with a public changelog; the HTML is roughly 170 KB and each version has a stable dated URL.

Methodology Notes

Normative specification, not an empirical study. Version dated 2026-08-18 per the page title and the changelog in the openai/model_spec GitHub repository; the December 2025 version (2025-12-18) sits between it and the held October 2025 version. Fetched directly from model-spec.openai.com on 2026-09-15 (the root redirects to the dated page); the accompanying openai.com blog posts are not reachable from this workspace. Applies to OpenAI's deployed models and ChatGPT.