OpenAI Rewrites Its Model Spec Around Five Core Values

When models were just chatbots, rules could be granular: what to say, what not to say, which tone to use. Once models write code, send email and operate systems on people's behalf, the rulebook can never be finished. The latest version of OpenAI's Model Spec reflects exactly that shift, from listing rules to articulating values.

On October 29, OpenAI updated the Model Spec, its public description of how it wants its models to behave. The new version is reorganized around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.

Five values, one foundation

Faithfulness to human intent is the foundation. OpenAI says faithfulness means going beyond following instructions literally to deeply understanding what OpenAI, developers and users intend and value, and acting wisely when that understanding is incomplete. It also means recognizing that people's beliefs change, and helping them reflect and grow in ways they would endorse, without quietly steering them or overriding their choices.

The other four make longer-term considerations explicit. Minimizing harm means respecting firm safety boundaries and accounting for foreseeable consequences. Honesty and transparency mean being truthful about what the model knows, does and is unsure of. Facilitating human agency means strengthening people's ability to understand, choose and act for themselves. Care means attending to the circumstances of those affected, following relevant social and professional norms, and doing work people can rely on.

The framework borrows from the assistance-games framing in alignment research: what people value comes first, and instructions are usually the clearest signal of those values. That is a meaningful conceptual turn. The model is defined less as an obedient assistant and more as an agent that understands what you're after.

What was added, what was cut

New material includes a subsection on considering context carefully, explaining how conversations, memories and other evidence should inform the model's read of people's intentions and preferences. There are new defaults on supporting human agency and following social norms, and expanded guidance on controlling and communicating side effects. Almost all of it points at agentic use: the more a model does, the more it must weigh effects on third parties.

Meanwhile OpenAI retired most of the detailed style guidance, arguing that style should adapt to preferences across people, languages and cultures. Stronger defaults remain, such as approaching questions objectively: users can explicitly ask for a perspective, but the model shouldn't infer one just from what it thinks they want to hear. That line is aimed squarely at sycophantic AI.

One small change is symbolic: the Spec now refers to the model as "GPT" rather than "the assistant," reflecting its widening range of uses and deployments.

The spec runs ahead of the model

OpenAI concedes the Spec describes behavior it is working toward, and its models don't fully reflect it yet. It is also exploring training methods meant to help models understand why the principles matter, to do the right things for the right reasons.

Across the industry, this version moves the Spec closer to the constitution approach Anthropic has long championed: governing increasingly autonomous models with values rather than exhaustive rules. The catch is that the more abstract the values, the more room for interpretation, and the harder it is for outsiders to verify compliance. Publishing the document and inviting feedback trades transparency for trust. What comes next is verifiable evidence that model behavior actually matches the spec.