OpenAI Rebuilds Its Model Spec Around Five Values and Drops Most Style Rules
The October update to OpenAI's public rulebook for model behavior adds guidance on context, oversight and human agency, and now calls the model GPT instead of the assistant.
Published
Last updated / Fact-checked on
OpenAI published an updated Model Spec on October 29, 2026, reorganizing its public description of how its models should behave around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision adds guidance on context, oversight and agency, changes some default behaviors and retires most of the document's detailed style rules.
Image: Illustrative photo of a laptop and a stack of books. Kari Shea (Unsplash) / CC0 1.0, via Wikimedia Commons
The five values
OpenAI says the values pull together commitments already spread across the Spec rather than introducing new ones. It also says they are not meant to capture every human value its models should respect.
Faithfulness to human intent is the foundation. OpenAI defines it as going beyond literal instruction-following to understand what OpenAI, developers and users intend, and acting wisely when that understanding is incomplete. The other four values, as the company describes them:
- Minimizing harm: respecting firm safety boundaries and accounting for the foreseeable consequences of the model's actions.
- Honesty and transparency: being truthful about what the model knows, does and is uncertain about.
- 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.
The Spec keeps its layered structure. OpenAI sets boundaries, developers direct applications within them, and users direct the model's work within those constraints. Outside the firm boundaries, most guidance takes the form of defaults that users and developers can override.
What is new
A new section on faithfulness to human intent includes an updated "chain of command," which explains which instructions apply and how conflicts are resolved. OpenAI says the approach draws on the "assistance games" framing from alignment research.
A new subsection, "consider context carefully," covers how conversations, memories and other context should shape the model's understanding of what people want. OpenAI's example: when helping someone compare job offers, the model might draw on priorities the person shared earlier, such as predictable hours, while making those assumptions visible and easy to correct.
The update also adds defaults on supporting human agency and following social norms. It expands existing guidance on controlling and communicating side effects, meaning the effects of a model's work on people beyond the user.
What is retired
Most of the detailed style guidance is gone. OpenAI says style should adapt to what people prefer, in aggregate, within languages and cultures, and for each individual, using feedback and context as evidence.
Some stronger defaults stay. The Spec still asks the model to approach questions objectively. Users and developers can ask for a particular perspective, but the model should not infer that preference just because it thinks the person wants to hear it.
The document now refers to the model as "GPT" rather than "the assistant," which OpenAI attributes to the growing range of products and settings where its models run.
The caveat
OpenAI says the Spec describes behavior it is working toward and that "our models do not yet fully reflect it." The company says it is testing new training methods meant to help models learn the underlying values, not just follow rules, and will report results later. It is also inviting public feedback on the values, trade-offs and wording.
Why it matters
As models take more actions on their own, fixed rules cover less of what they encounter. OpenAI is betting that teaching values, plus a clear order of whose instructions count, will produce better judgment in situations nobody wrote a rule for.
For developers, the practical changes are the updated chain of command and the guidance on context and memory, which affect how models interpret instructions inside apps and agents. Anyone building on OpenAI models should reread those sections and test edge cases in their own products. Anyone who disagrees with the trade-offs can send feedback through the channel OpenAI opened with the update.
More coverage of AI safety and policy is in our Artificial Intelligence hub.
How we reported this
We checked every claim against OpenAI's announcement post. We cut the draft's description of specific oversight requirements, such as shutdown and replacement, because the announcement does not detail them, and we did not separately verify the full Spec text.
Sources
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.