OpenAI Overhauls Its Model Spec Around Five Values and Now Calls Its Model “GPT”
OpenAI has published a new version of its Model Spec, the public document outlining the expected behavior of its models. It is now organized around five core values, led by fidelity to human intent, adds rules covering context, memory, human oversight, and side effects, and drops most of its detailed style guidelines.
OpenAI published a new version of its Model Spec on Thursday, October 29, the public document in which the company describes how it wants its models to behave. Announced around 6:05 p.m. Paris time by researcher Jason Wolfe, and subsequently reposted by OpenAI’s account, this revision goes beyond merely adding rules: it reorganizes the entire document around five core values, based on the premise that more autonomous models must understand what people want rather than simply following their instructions verbatim.
Five values instead of a list of rules
The five chosen values are:
- Fidelity to human intent, framed as the foundation: understanding what OpenAI, developers, and users want, and exercising sound judgment when that understanding is incomplete;
- Harm reduction: adhering to firm safety boundaries and considering the foreseeable consequences of the model’s actions;
- Honesty and transparency regarding what the model knows, does, and does not know;
- Empowering human agency: helping people understand, choose, and act on their own;
- Care for people who are affected, respect for social and professional norms, and delivering reliable work.
The instruction hierarchy remains in place: OpenAI establishes the hard boundaries, developers shape their applications within them, and users direct the model’s work within that framework. Outside the strict boundaries, the Spec remains primarily a set of default behaviors that developers and users can modify.
What was added, what was dropped
A new section, “Consider context carefully,” explains how conversations, memories, and other contextual elements should inform the understanding of user intent in a relevant, proportionate way that respects privacy. OpenAI offers an example: when comparing job offers, the model can remember that the user values predictable hours, but it must make this assumption explicit and easy to correct.
The text also adds principles on supporting human oversight and correction, and strengthens instructions to control and communicate side effects of the model’s actions—with additional specifics on permissions, a crucial detail for agents acting within software or across the web.
Conversely, OpenAI drops most of the detailed style guidelines: style should adapt to user preferences, languages, and cultures. One strong rule remains: addressing questions objectively, without adopting a point of view simply because the model assumes it is what the user wants to hear. Finally, the document now refers to “GPT” rather than “the assistant.”
A statement of intent, not a guarantee
OpenAI acknowledges it directly: the Model Spec describes intended behavior, and its current models do not yet fully reflect it. The company indicates it is working on new training methods to help models understand why these principles matter, though it has not published any results for now. It invites public feedback on the text.
Our take
Shifting from a list of rules to a foundation of values is a gamble: rather than anticipating every scenario, the hope is that the model can reason through unexpected situations. This aligns with the rise of agents, which make decisions no human has scripted in advance. Yet it also proves harder to evaluate from the outside: a precise rule can be directly verified, whereas a value like “fidelity to intent” requires interpretation. The candid admission that models do not yet fully align with the text serves as a reminder that this is a roadmap—one that will be judged by how future models actually behave.
Sources
- OpenAI — An updated Model Spec for more capable and beneficial AI Primary source
- OpenAI — Model Spec, October 29, 2026 version Primary source
- OpenAI (GitHub) — Model Spec changelog Primary source
- OpenAI Help Center — Model Release Notes Primary source
- Jason Wolfe (X) — Model Spec update announcement Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
