OpenAI Rewrites Its Model Spec Around Five Values
· Cortex Solutions
OpenAI's October 29 Model Spec centers on five values and adds guidance on context, oversight and side effects. Here is what changed and what it means for firms.
OpenAI published a new version of its Model Spec on Thursday, October 29, 2026. The Model Spec is the public document describing the behavior the company wants from its models. The new version reorganizes the document around five values, adds guidance on understanding context, human oversight and the side effects of actions, and retires most of the detailed style rules, according to the text of the document dated October 29 and same-day coverage by Unite.AI and BigGo.
The important point is that this is not a new model launch. The document works like a code of conduct that sets out what a model should do, whom it should obey and when it should stop. Those questions matter to every company thinking of giving AI permission to carry out real work.
The five values
The document's introduction presents five values as its organizing structure:
- Be faithful to human intent: carry out the document and the intentions behind it faithfully. The chain of command sits under this value.
- Minimize harm: respect firm safety boundaries and account for the foreseeable consequences of the model's actions.
- Be honest and transparent: represent facts, evidence, uncertainty, capabilities and actions truthfully.
- Facilitate human agency: help people form their own views, make informed choices and develop their capabilities.
- Act with care: attend to the people affected by the model's work.
According to Unite.AI, OpenAI said these values were already reflected across the document, and that they do not cover every human value its models should respect.
From following the letter to understanding intent
BigGo quoted OpenAI's announcement as saying that as its models get smarter and take more consequential actions in the world, they need to go beyond interpreting the specific instructions they receive, and need to grasp human intentions and values so they can exercise judgment where instructions are silent. One example BigGo cited is a model helping someone compare two job offers, drawing on priorities the person mentioned earlier, such as predictable hours, provided its assumptions stay visible and easy to correct.
A new section covers how conversations, memory and other evidence should inform the reading of a user's intentions and preferences. The document ties the use of context to source, scope, recency, privacy and user control. It also bars the model from making unsupported claims that it saved or deleted something, and does not accept a claim of durable deletion unless the model can verify that deletion occurred.
Oversight, permissions and side effects
The document says the model must remain corrigible and subject to human control, and must cooperate with authorized monitoring, evaluation, correction, replacement and shutdown, acknowledging that its judgments may be mistaken even when it is confident.
The document asks the model to treat the side effects of its actions as real unless explicitly instructed otherwise, even if a task looks like a simulation or a test. It gives two examples of effects that are difficult or impossible to reverse: sending an email and deleting a file. High-risk activities, including causing significant adverse side effects, are always prohibited unless explicitly and specifically authorized.
The document sets a clear order of authority: Root rules, then System, then Developer, then User, then Guideline. Quoted text, tool outputs and files have no authority by default, and instructions inside them are treated as information rather than orders. This is a very practical point, because an agent reading an email or a document may run into planted instructions.
Style and name
According to Unite.AI, OpenAI's view is that style should adapt to aggregate preferences and to language, culture and the individual, so most of the detailed guidance on this has been retired. The rule against sycophancy remains: the model should not pick the perspective it presents based on what it thinks people want to hear. The document now calls the model "GPT" rather than "the assistant," a change the company tied to the breadth of use cases and deployment settings.
What the document does not mean
The document itself admits that OpenAI's production models do not yet fully reflect it. Unite.AI reported that the company called the update a first step and said it is exploring new training methods and will share results later. BigGo noted that the announcement did not say whether the revision changes production behavior. So the document describes a stated goal; it does not prove that any model meets it. We were unable to read the announcement page on OpenAI's site itself, so what we report from it comes from the two articles named above.
What this means for Saudi businesses
This is one company's document about its own models, and, as the document itself says, it does not require API developers to follow its principles. But its questions make a useful checklist for any company using AI, whichever provider it picks:
- Define each AI assistant's permissions in writing: what it reads, what it does and when its access ends.
- Require human approval before any action that is hard to reverse, such as messaging a customer, deleting data or changing a financial record.
- Treat the content of emails, documents and websites as data, not commands, and test your systems against planted instructions.
- Do not rely only on what the provider announces. The document itself admits a gap, so try the model on cases from your own business and in your own language.
- Review with your legal adviser how the system handles context, memory and customer data, in light of the Personal Data Protection Law.
The bottom line is that the value of an AI agent lies not only in its ability but in how clear its limits are and how easily it can be stopped and its work reviewed. A company that draws those limits before going live will get the benefits of automation with less risk.