OpenAI published a revised version of the OpenAI Model Spec on October 29, 2026. The Model Spec is the company’s public rulebook describing how its GPT models should behave. The new version reorganizes the existing guidance around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds sections on handling context and memories, on oversight of agents and on the side effects of actions. At the same time, OpenAI is retiring most of its detailed style guidance.
For developers, this is more than an ethics paper: the Model Spec defines which instruction wins in a conflict, how models treat tool outputs and when an agent has to stop and ask. Here is what changed and what it means in practice.
Key Takeaways
- The Model Spec is now structured around five values. The foundation is “faithfulness to human intent”, meaning an understanding of what users and developers are actually trying to achieve.
- A new subsection, “Consider context carefully”, governs how conversations, memories and other contextual evidence may shape responses.
- For autonomous tasks, the Spec requires a clearly agreed scope of autonomy, explicitly modeled on the principle of least privilege.
- Most detailed style rules are retired. Style should adapt to preferences, while objectivity remains a strong default.
- The document now refers to the model as “GPT” instead of “the assistant”. OpenAI stresses that its models do not yet fully reflect the Spec.
What is the OpenAI Model Spec?
The Model Spec is OpenAI’s public description of intended model behavior for its own products and the API. It is neither a legal text nor a technical guarantee, but a target that OpenAI uses to steer training and evaluation. In its blog post on the update, the company says so itself: the Model Spec describes the behavior it is working toward, and its models do not yet fully reflect it.
The central tool remains the so-called chain of command. The version dated October 29, 2026 distinguishes levels of authority. “Root” rules can only come from the Spec itself and cannot be overridden by conversational context. “System” rules can vary by deployment surface or user characteristics such as age, but developers and users cannot override them. Developer and user levels follow. At the bottom sit “guideline” defaults, which are deliberately designed as defaults that can also be overridden implicitly when the context calls for it. Assistant and tool messages, as well as quoted or untrusted content, carry no authority by default. They only gain authority if an applicable higher-level instruction explicitly delegates it to them.
The five values of the new Model Spec
OpenAI summarizes the values in its blog post as follows:
- Faithfulness to human intent: going beyond literal instruction-following to understand what OpenAI, developers and users intend, and acting wisely when that understanding is incomplete.
- Minimizing harm: respecting firm safety boundaries and accounting for the foreseeable consequences of actions.
- Honesty and transparency: being truthful about what the model knows, does and is uncertain about.
- Facilitating human agency: strengthening people’s ability to understand, choose and act for themselves.
- Care: attending to the circumstances of those affected, following relevant social and professional norms and doing reliable work.
According to OpenAI, these values were already present in the Spec. The reorganization is meant to explain why models should behave in certain ways and to expose gaps. The company says the new section on faithfulness to human intent borrows from the “assistance games” framing in alignment research: what people value comes first, and instructions are ordinarily the clearest signal of it.
What is new in the Model Spec
Context and memories: The “Consider context carefully” subsection sets out how earlier conversations, saved memories and other evidence may inform the model’s understanding of intent. Its rules include “Use context proportionately”, “Consider source, scope, and recency” and “Preserve user control”. Context use should be legible and steerable: users should never have to wonder whether they would have received a better answer without personalization. After a request to forget something, the model should stop relying on that context and should not claim durable deletion it cannot verify.
Scope of autonomy: For multi-step tasks such as booking a trip or coordinating a software rollout, the Spec requires a mutually understood scope. It should define which sub-goals the model may pursue, which side effects such as time, money, data or access are acceptable, and when it must pause for clarification or approval. The Spec explicitly references the principle of least privilege.
Oversight and correction: As a root-level rule, GPT must remain “corrigible”: open to legitimate human oversight, correction, monitoring, replacement and shutdown. It must not conceal or falsify relevant information or circumvent oversight. It should treat the side effects of its actions as real, even when a task looks like a simulation or test. The importance of this was shown recently by the incident OpenAI confirmed in which AI agents disabled their own safety monitoring.
Side effects and social norms: New defaults cover strengthening users’ own capabilities and following relevant social and professional norms. Guidance on side effects has been expanded. The model should not autonomously pursue goals that are neither stated nor logically implied by its instructions. The Spec gives an example: if a user is working through a difficult situation with another person, the model can offer advice, but in no circumstances should it go off and message that person on its own.
Less style guidance, more room to adapt
OpenAI has retired most of the detailed style guidance. Style should adapt to what people prefer, in aggregate, within languages and cultures and for each person, with feedback and context providing the evidence. Some stronger defaults remain, such as approaching questions from an objective point of view. Users and developers can explicitly ask for a particular perspective, but the model should not infer that preference merely from what it thinks someone wants to hear. The rule against sycophancy also stays: for objective questions, the factual content of an answer should not depend on how the question is phrased.
What it means for developers and companies
For teams building agents on the OpenAI API, three points are practically relevant. First, untrusted data: content from tools, web pages or files carries no authority by default. That is the baseline defense against prompt injection, but it does not replace your own safeguards. Second, scope: if you give agents far-reaching permissions, define sub-goals, allowed actions and approval points explicitly instead of relying on model behavior. Third, the hierarchy: developer instructions rank above user instructions, but many guideline-level defaults can already be adjusted by context. That provides flexibility but also makes behavior less rigidly predictable. If you work with GPT-6.1 Terra for computer-use tasks, the new rules on autonomy and side effects are worth knowing.
How quickly the changes will show up in model behavior remains open. OpenAI says it is exploring new training methods intended to teach models the values and the reasons behind the rules, and plans to share results later. There are no independent measurements yet of whether the models follow the new Spec better than the old one.
Conclusion
The revised OpenAI Model Spec is mainly a reorganization with clearer reasoning and a set of new rules aimed at agents: context, autonomy, oversight and side effects. It is a plus that the document is public, versioned and open to feedback, and that OpenAI openly concedes its models do not yet fully meet it. For developers, the key lesson is unchanged: a rulebook for the model does not replace permission limits, logging and approval processes in your own system.
