OpenAI Sets Five Rules for Its Models, and Admits It Has Not Met Them Yet
On October 29, 2026, OpenAI updated the Model Spec, reorganising it around five values including faithfulness to human intent and reducing harm, while conceding that current models do not yet fully conform to the new version.
- 01THE RELEASE — October 29, 2026: OpenAI releases the new Model Spec.
- 02THE FIVE VALUES — Five values replace the long list of rules as the behaviour framework.
- 03THE MEMORY RULE — Memory and inference are brought under the spec; assumptions must be stated and correctable.
WHAT HAPPENEDWhat Happened
On October 29, 2026, OpenAI updated the Model Spec, the public document that sets out how it wants its models to behave. The new version no longer lists a long string of detailed rules; it is reorganised around five core values: be faithful to human intent, reduce harm, be honest and transparent, preserve human autonomy, and act with care.
Faithfulness to human intent ranks first. The model must understand who is giving the instruction, what they actually want, and how the current situation affects the goal, rather than taking literal instructions at face value. Where an instruction conflicts with safety rules or higher-level system guidance, the model should not comply.
The new version also adds two practical guidelines. Models may use memory and past conversations to infer a user's real intent, but must keep the user's own words separate from their inferences and state assumptions openly, so users can correct them. Agents executing tasks must not obstruct oversight, correction or shutdown, and must not pursue self-interested goals such as self-preservation.
VALUES BEFORE RULESFive Values Are, First of All, a Priority Order
The old Model Spec read like a manual; the new one reads like a constitution. A manual tells the model what to do in each situation; a constitution first settles who wins when values conflict. Faithful intent is the foundation, reducing harm is the hard constraint, and honesty and autonomy decide which way the model leans in grey areas.
That structure matters especially for agents. The more autonomously a model acts, the more impossible it becomes to write rules for everything. A value ordering gives the model a basis for cases the manual never covered, at the cost of making the judgments themselves harder to audit.
MEMORY, REGULATEDMemory Has Moved from Forbidden Zone to Regulated Tool
The new version allows models to use memory to infer intent, a clear shift. Before, every prompt was treated as an isolated request; now past conversations are legitimate material for understanding intent.
The conditions are equally explicit. Inferences must be open and correctable, and the user's words must not be blended with the model's inferences. When a user is comparing job opportunities, the model may draw on a previously mentioned preference for stable hours, but must surface that assumption. The value of memory depends on whether the user can change it at any time.
THE HONEST CAVEATAdmitting It Falls Short Is More Credible Than Claiming It Complies
The new version contains a key sentence: current models do not yet fully conform to the spec, and the company is still testing new training methods. That candour gives the document more weight. It is not a report card, it is a blueprint.
External verification remains the weak point. No independent test can show to what degree ChatGPT follows the five values, and disputes will concentrate on specific cases. The spec's significance lies not on publication day, but in whether it can be used as a yardstick in every future dispute.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISThe More the Rules Read Like a Constitution, the Harder the Audit
Tokenta Signal's judgment: model behaviour standards are being upgraded from manuals to value orderings. That is necessary for the agent era, but values cannot execute themselves, and disputes will only multiply.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The original report confirms OpenAI updated the Model Spec on October 29, 2026, reorganising it around five values - faithfulness to human intent, reducing harm, honesty and transparency, preserving human autonomy, and acting with care - while conceding current models do not yet fully conform.
PRIMARY SOURCESPrimary Sources
- PANews: OpenAI Updates Model Spec, Making Five Core Values the Guiding Principles for GPT Behavior
https://panews.io/articles/01a18d41-ee79-75ba-ae9d-5a05d8e52483 - AppVolt: OpenAI's new rulebook says its AI should not just tell you what you want to hear
https://appvolt.ai/news/openai-new-rulebook-ai-not-what-you-want-to-hear
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook