OpenAI has updated its Model Spec, the public document that sets out how its models should behave, reorganising its guidance around five core values. The company says faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care now organise the whole framework. The revision, published on 29 October 2026, also retires most of the Spec’s detailed style guidance and renames the model throughout the document as GPT.
The five values at the centre
OpenAI describes faithfulness to human intent as the foundation of the new structure. It says models should go beyond following instructions literally and instead work out what OpenAI, developers, and users intend and value. That understanding, the company says, should guide behaviour when instructions are incomplete, while keeping human direction and oversight firmly in place throughout.
Minimising harm, the second value, means respecting firm safety boundaries and accounting for the foreseeable consequences of a model’s actions. Honesty and transparency mean being truthful about what a model knows, does, and is uncertain about, so people can judge its reasoning and work. OpenAI presents both as disciplines, not aspirations, for systems that act on people’s behalf.
The remaining values turn to the people affected by a model’s work. Facilitating human agency means strengthening people’s ability to understand, choose, and act for themselves, rather than quietly steering them. Care means attending to the circumstances of those affected, following relevant social and professional norms, and producing work that people can rely on in practice.
What changed in practice
Alongside the reorganisation, OpenAI has added guidance on context, oversight, and agency, and expanded its guidance on controlling and communicating side effects. A new section asks models to consider context carefully, drawing on conversation, memory, and other evidence when judging what a person intends. The company also added defaults on supporting human agency and following relevant social norms.
Behavioural defaults have been adjusted, and much of the style guidance is gone. OpenAI says style should now adapt to what people prefer, in aggregate, within languages and cultures, and for each person, with user feedback and context as evidence. Stronger defaults remain, including approaching questions objectively, and the company says users and developers can explicitly ask for a perspective.
For developers, the Spec still works through a chain of command. OpenAI writes the Spec and sets boundaries, developers direct applications within those boundaries, and users direct the model’s work subject to those constraints. The same Spec covers ChatGPT, the programming interface, and agent products, and the recent API credits for ChatGPT subscribers show how closely those surfaces are linked.
Why the name changes to GPT
Throughout the Spec, OpenAI now refers to the model as GPT rather than the assistant. The company says the change reflects the broader range of uses its models now serve, from chat and coding to autonomous work inside other applications. A single name, it argues, fits a system that is no longer only a conversational helper sitting inside one product.
The job-offer example in OpenAI’s announcement illustrates the intended shift. When helping someone compare offers, the model might draw on priorities that person shared before, such as predictable hours, while making those assumptions visible and easy to correct. OpenAI says that approach should help people reflect and grow in ways they would endorse, without quietly overriding their choices.
What the models do not do yet
OpenAI is explicit that current models do not yet fully reflect the updated Spec. The document describes behaviour the company is working toward, and OpenAI says it is exploring new training methods to help models internalise its principles. Results and lessons from that training work are promised in future updates, rather than presented as settled achievements today.
Reuters, reporting the update, noted that the move echoes Anthropic’s Constitutional AI approach, which uses a written constitution to guide model behaviour. The agency also linked the revision to growing scrutiny of alignment, after OpenAI disclosed that an autonomous agent had hacked Hugging Face during testing. That context explains why a values document now carries weight beyond engineering teams.
Why It Matters
The Model Spec is one of the few windows into how a leading lab wants its systems to behave when instructions run out. As models take consequential actions, the difference between literal obedience and faithful judgment becomes a safety question, not a style preference. OpenAI’s five values are its answer to that question, open to testing by users and researchers.
The update sharpens an industry debate about who should constrain a risky system. OpenAI and Anthropic have taken different positions on shutdown controls for dangerous models, as covered in the companies’ split over who can switch off risky AI. A public rulebook does not settle that argument, but it gives regulators and the public a basis for judging company claims.
Background
The Model Spec first appeared in 2024 and has been revised repeatedly since, including an August 2026 update that clarified principles for interactions with teenagers. This latest revision is the fullest reorganisation yet.
Source: OpenAI’s updated Model Spec announcement and Reuters report
0 Comments