OpenAI updated its Model Spec on 29 October 2026, reorganising its public description of how it wants models to behave around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The revision gives developers and users a clearer account of how OpenAI wants models to handle instructions, context and decisions that instructions do not fully cover. It also retires much of the document’s detailed guidance on style. OpenAI says the Spec describes behaviour it is working towards, rather than behaviour its models already exhibit in full.
Key points
- Faithfulness to human intent is the foundation of OpenAI’s revised guidance; four other values address harm, truthfulness, agency and care.
- An updated chain of command governs conflicting instructions, while new guidance asks models to consider conversational context.
- OpenAI has retired most detailed style rules and says its models do not yet fully reflect the revised Spec.
Faithfulness leads OpenAI’s five values
OpenAI places faithfulness to human intent at the centre of the revision. In its account, OpenAI writes the rules and teaches models to follow them, while people building with the models decide their uses and users choose tasks subject to those rules. Faithfulness calls for more than taking each request literally: the model should try to understand what the instruction is meant to accomplish, including when that understanding is incomplete.
The other four values give that approach further obligations. OpenAI says harm minimisation requires attention to firm safety boundaries and foreseeable consequences. Honesty and transparency concern what a model knows, does and is uncertain about; human agency concerns people’s ability to understand and make their own choices. Care asks the model to take account of affected people’s circumstances and relevant social and professional norms. OpenAI says these five commitments are not intended to encompass every human value a model should respect.
The distinction matters when a request has more than one plausible interpretation. OpenAI gives the example of comparing job offers: a model might remember that predictable hours were important in an earlier conversation, but should expose that assumption so it can be corrected. The new guidance on considering context extends beyond the latest instruction to conversations, memories and other evidence of a person’s preferences, while allowing for those preferences to change.
Comparing job offers could therefore draw on a previously stated preference for predictable hours, provided the assumption were made visible and easy to correct. It would leave the choice with the person making it, rather than quietly treating an earlier preference as a permanent instruction.
The Model Spec’s updated chain of command
The new faithfulness section incorporates an updated chain of command, OpenAI says, to determine which instructions apply when they conflict. The practical idea resembles standing rules for a task: a new request can direct the work, but it does not erase the rules under which that work is done. OpenAI describes much of the Spec as defaults that developers and users can override, alongside firmer boundaries intended to protect people and society.
Unite.AI reported that the revised document places root-level instructions above system, developer and user instructions. Its account says quoted material and tool messages carry no authority by default, while guideline-level instructions can yield when context clearly calls for it. Those distinctions matter when a model encounters instructions inside material it has been asked to process: the material can inform the task without automatically gaining the authority to redirect it.
The same publication reported that rules for autonomous work require a mutually understood scope covering permitted sub-goals, acceptable side effects and points at which the model must seek clarification or approval. OpenAI’s announcement also says it has expanded guidance on controlling and communicating side effects. For developers building applications in which a model can take actions, these provisions specify the boundaries OpenAI wants the model to observe as well as the work a user asks it to perform.
OpenAI retires detailed style guidance
OpenAI has removed most of the Spec’s detailed advice on style, saying presentation should adapt to people’s preferences as indicated by feedback and context, including differences across languages and cultures. It has kept stronger defaults, including an objective approach to questions. Developers and users can ask for a particular perspective, but OpenAI says a model should not adopt one merely because it guesses that perspective will be welcome.
The revision adds defaults on supporting human agency and following relevant social norms. OpenAI also now calls the model “GPT” throughout the Spec rather than “the assistant”, a wording change it attributes to the wider range of settings in which its models are used. Together, the changes give developers room to direct an application’s behaviour while retaining OpenAI’s stated boundaries and its expectations about how the model handles other people affected by its work.
Alongside the document update, OpenAI says it has been exploring training methods intended to help models learn the values and apply the Spec’s principles in unfamiliar situations. It says subsequent updates will cover its findings and what it has learned from that training. For now, the company says its models do not yet fully reflect the behaviour described in the Spec.
OpenAI is inviting public feedback on the revised values, the trade-offs between them and the document’s specific wording.
