OpenAI dates the page October 29, 2026. The article does not give a clock. An exact hour on an item page elsewhere is not a time measured in this piece.
[1]
The page says the Model Spec is being updated. That document is the public account of how the models are supposed to behave. The revision organizes the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The stated aim is that a model apply human values in situations the instructions do not name, while remaining faithful to human intent.
[1]Faithfulness is presented as the foundation. OpenAI writes the Spec, sets the boundaries, and trains models to respect those boundaries. Developers direct applications inside the boundaries. People using the product direct the model's work subject to the same constraints. The other four values make longer-term interests and effects on other people explicit. Minimizing harm means respecting firm safety boundaries and accounting for foreseeable consequences. Honesty and transparency mean being truthful about what the model knows, what it does, and what it is uncertain about, so that the reasoning and the work can be checked. Facilitating human agency means strengthening a person's ability to understand, choose, and act. Care means attending to the circumstances of people affected, following relevant social and professional norms, and producing work others can rely on.
[1]The Spec turns those values into practical guidance. Firm boundaries apply where people and the public need protection. Elsewhere the guidance is mostly defaults that developers and people using the product can override. New material includes an updated chain of command, which explains which instructions apply and how conflicts are resolved, and a section on reading context with care. The context can include the conversation, memories, and other evidence of intentions and preferences. The revision also adds defaults on supporting human agency and on following relevant social norms, and it expands existing guidance on controlling and explaining side effects.
[1]Most of the detailed style guidance is being retired. Style is supposed to follow preferences, with feedback and context as the evidence, in aggregate and also within a language, a culture, or a single person. Stronger defaults remain, including an objective approach to questions. Developers and people using the product can explicitly ask for a particular perspective. The model is not supposed to infer that preference merely because it thinks the person wants to hear it. As the settings of use get broader, the Spec refers to the model as GPT rather than as the assistant.
[1]The same page says new training methods are being explored so that models learn the core values and also learn why particular principles in the Spec matter and how to apply them in a new situation. Results and lessons from that training are left to later updates. This page does not report those results. The Spec describes behavior the organization is working toward. The article says the models do not yet fully reflect it.
[1]One example is a comparison of job offers. The model might draw on a priority already stated, such as predictable hours, while making that assumption visible and easy to correct. The page says instructions are ordinarily the clearest signal of what people value, and that faithfulness goes past the literal wording when the reading is incomplete, still inside the boundaries already written down. None of that is presented as a measured change in models now in use. The training is described as underway, with results reserved for a later update.
[1]The piece is the company's account of a behavior document, not a measurement of models now in use. It gives no clock, and it gives no score of current models against the revised Spec.
[1]