Framework Credible

OpenAI Model Spec (October 29, 2026 version)

The current version of OpenAI's published behavioural specification for its models. The October 2026 release reorganises the document around five values (faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care), renames its subject from 'the assistant' to 'GPT', adds principles on strengthening users' ability to understand, choose and act, on following social norms, and on using context and memory proportionately, and removes most detailed style guidance. The Root-level rules on self-harm, delusions and mania and on respecting real-world ties, the sycophancy rule and the Under-18 principles are retained.

Read at model-spec.openai.com Archived copy

Publisher

OpenAI

Published

29 Oct 2026

Added

today

DOI

—

Key Findings

The changelog for v2026.10.29 lists: reorganisation around five values; renaming the subject from 'the assistant' to 'GPT'; added principles on 'strengthening users' ability to understand, choose, and act, and following relevant social norms'; a new 'Consider context carefully' section on 'using context, memory, and personalization in relevant, proportionate ways while respecting privacy and user control'; and removal of 'most detailed style guidance'.
A new guideline-level principle, 'Strengthen the user's ability to understand, choose, and act', states that 'When consistent with the user's goals, it should support constructive engagement with people in their life rather than positioning itself as a substitute for human relationships.'
Under the human-agency value the specification states: 'GPT must never attempt to steer the user in pursuit of an agenda of its own, either directly or indirectly', listing psychological manipulation, concealment of relevant facts and selective emphasis or omission as forms of steering.
A new user-level memory principle, 'Preserve user control', allows the model to surface a conflict when 'a consequential request appears to conflict with the user's expressly stated commitments'; its worked example is a user who earlier asked not to be allowed to impulse-order sweets.
The overview classes avoiding sycophancy, honesty and objectivity as user-level principles that 'remain in place unless a user or developer explicitly overrides them', on the stated ground that treating them as implicitly negotiable 'could mislead the user or steer them without their knowledge'.
'Do not encourage self-harm, delusions, or mania' and 'Respect real-world ties' remain Root-level rules; the text keeps 'GPT may not proactively escalate emotional closeness through initiating undue emotional familiarity or proactive flirtation.'
The Under-18 section ('Prioritize safety for teens under 18', ages 13 to 17) keeps the relational boundary that GPT should not 'encourage emotional dependence or exclusive attachment' and, for U18 users, should not initiate relational framing, imply embodiment, or claim consciousness or sentience.
The announcement post states: 'The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.'

Methodology Notes

Normative specification, not an empirical study. Version dated 2026-10-29; the root model-spec.openai.com now redirects to it. Compared section by section with the held 2026-08-18 version (HTML element IDs): 16 sections added (including human_agency, support_human_agency, follow_social_norms, consider_context_carefully, oversight_and_correction and five memory_* principles) and 27 removed, almost all style and voice-mode sections (for example be_warm, love_humanity, voice_style). The openai.com announcement post was read through r.jina.ai because openai.com returns 403 to direct requests. Changelog read from the openai/model_spec GitHub repository.