OpenAI rewrites its model behaviour rules around five core values

OpenAI released a major revision of its Model Spec on 29 October 2026, setting out how its GPT systems should behave as they take on more autonomous tasks for businesses and individual users.

The Model Spec is the public framework OpenAI uses to train and steer its models. The new edition reorganises those rules around five core values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. In the document, OpenAI also officially refers to the model as GPT rather than the assistant.

OpenAI explained that as systems carry out longer tasks and make decisions in software environments, they can no longer rely purely on word-for-word instructions. The updated rules establish a clear order of authority: OpenAI sets the baseline safety boundaries, software developers configure the operational bounds for their applications, and end users direct the specific work within those limits.

Less rigid style, more focus on agent actions

The update removes most of the detailed writing-style rules that were included in previous versions. Instead of enforcing set lengths or boilerplate phrasing, OpenAI wants GPT to adapt tone and presentation to the specific instructions given by the user, while keeping an objective tone as a default.

The document also adds practical guidance for autonomous agents. Under the revised framework:

• Models must remain open to correction and keep their actions strictly within the scope agreed with the user.
• Models should treat instructions inside untrusted third-party content, such as web pages, emails or file uploads, as data rather than commands.
• Models must separate what a user explicitly stated from inferences drawn from context or memory.

OpenAI noted that its current systems do not yet fully reflect all the behaviours in the updated specification, describing the document as the target for ongoing training and model updates.

What this means for Australian businesses

For Australian companies using ChatGPT or building on OpenAI application programming interfaces, this update signals where software defaults are heading.

• Check your prompt instructions. With OpenAI stripping away its built-in style constraints, your system prompts will carry more weight in determining the exact voice, formatting and length of your business outputs.
• Review permissions on automated agents. If your team is testing automated tools to handle customer correspondence, process incoming documents or triage support tickets, make sure those tools operate inside clearly defined task boundaries rather than open-ended access.
• Safeguard against prompt injection. The rules instruct models to ignore instructions embedded in outside text, but businesses that feed customer emails or supplier files directly into automated workflows should still keep validation checks in place before actions are taken.