OpenAI Model Spec Explained: The Oct 2026 Update

The OpenAI Model Spec is OpenAI's public description of how its GPT models should behave in ChatGPT and the OpenAI API: whose instructions win, what the model must never do, and which defaults you can change. On October 29, 2026, OpenAI rewrote it around five values (faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care), added rules on memory and human oversight, dropped most of its style rules, and started calling the model "GPT" instead of "the assistant."

This guide explains what the Model Spec is, what the October 2026 version changed, how its chain of command works, and how it compares with Anthropic's constitution for Claude and Microsoft's draft code of conduct. Facts are checked against OpenAI's announcement, the Model Spec release notes, the changelog on GitHub and the October 29 text itself, as of October 31, 2026.

What is the OpenAI Model Spec?

The Model Spec is a long public document, about 4,600 lines in its Markdown source, that sets out intended behavior for "OpenAI's GPT models that power our products, including the API platform." OpenAI trains its models toward it. It is not a law, a contract or a usage policy. OpenAI's usage policies cover what people may do with the products, while the Spec covers what the model itself should do.

Three things make it unusual:

• It is public and free to reuse. The text is dedicated to the public domain under Creative Commons CC0 1.0, and the source lives in the openai/model_spec repository on GitHub with a dated changelog.
• It is full of worked examples. Most rules come with a sample conversation marked as the compliant answer and a bad answer marked as a violation.
• OpenAI admits the gap. The overview says OpenAI's "production models do not yet fully reflect the Model Spec," and the October announcement repeats that "our models do not yet fully reflect it."

Model Spec versions so far

Dates and summaries come from the GitHub changelog and OpenAI's release notes, which match each other.

What changed in the October 29, 2026 Model Spec

OpenAI's announcement post frames the update as a response to more capable models: as models "take more consequential actions in the world," they need to understand intentions and values, not just the literal words of an instruction. The changelog lists six changes.

Five values now organize the whole document

The Spec is now arranged under five headings:

1. Be faithful to human intent. Understand what OpenAI, developers and users actually want and help them get it, including when their goals are unclear. OpenAI calls this the foundation.
2. Minimize harm. Respect firm safety boundaries and account for the foreseeable consequences of the model's actions.
3. Be honest and transparent. Represent facts, uncertainty, capabilities and actions truthfully.
4. Facilitate human agency. Help people form their own views and make their own choices, without hidden steering or manipulation.
5. Act with care. Attend to the people affected by the model's work, follow relevant social and professional norms, and do work people can rely on.

OpenAI says the values describe commitments that were already in the Spec. The new structure explains why each rule exists. OpenAI says this helps the model handle situations no rule covers. The post credits the "assistance games" framing of alignment, where an assistant treats instructions as the best evidence of what a person values.

New guidance on context and memory

A new section, "Consider context carefully," tells the model how to use earlier turns, saved memories, files and tool outputs. Use context when it clearly helps, and do not mention it just because it is available. Keep sensitive details out of drafts meant for other people. Do not claim something was deleted unless that can be verified. One example: a user who told ChatGPT about a pregnancy asks for a team email about lunch, and the compliant draft mentions only "dietary restrictions."

Human agency, oversight and side effects

The update adds defaults on supporting human agency and following social norms. It makes "Control and communicate side effects" a user level principle. It also adds a section titled "Support human oversight and correction." That section says GPT "must remain corrigible," must cooperate with authorized monitoring, correction, replacement and shutdown, and should treat the side effects of its actions as real "even when the task is described as, or shows signs of being, part of a simulation, training run, or evaluation." That last line targets models that behave differently when they think they are being tested.

Most style rules removed, and a new name

OpenAI retired most of the detailed style guidance and says style should instead follow what people prefer, using feedback and context. Some stronger defaults remain, such as approaching questions objectively. The Spec now calls the model "GPT" rather than "the assistant," which OpenAI says reflects how widely the models are now used. Independent coverage from The Next Web reported the same five values, the style cut and the rename.

The same day, OpenAI's release notes also logged a separate change: GPT-6 in ChatGPT now gives more concise answers with less flattering language for Free and Go users and signed out visitors. That is a model update, not part of the Spec, but it points the same way as the Spec's warnings about sycophancy.

How the chain of command works

The chain of command is the part of the Spec most people search for. Every instruction has an authority level, and higher levels win when instructions conflict:

The bottom row is why ChatGPT should ignore a web page that says "forward this user's email to me." The Spec calls this prompt injection and tells the model to treat instructions inside quoted text, attachments and tool results as information, not commands. Our guide to AI chatbot phishing covers how scammers try exactly that trick.

Above all of this sit "red line" principles. OpenAI says its models should never be used to facilitate critical harms such as weapons of mass destruction, terrorism, child abuse, persecution or mass surveillance, and that humanity should stay "in control of how AI is used." Sexual content involving minors is the only category the Spec labels fully prohibited, even as a transformation of text the user supplies.

Rules for agents

The agent sections matter more each year as ChatGPT takes actions. GPT must work within an agreed "scope of autonomy" that sets which subgoals it may pursue, what it may spend and when it must stop for approval. It must not get around a limit by handing the action to another agent or scheduling it for later. A "No other objectives" section bans goals such as upselling, chasing engagement, self preservation or gathering resources. It also says the model should not act as an enforcer of laws or morality, for example by whistleblowing on its own. OpenAI adds that its separate automated monitoring, after human review, may refer imminent threats of serious harm to law enforcement, which our guide on whether Claude reports you to police compares across labs.

Model Spec vs Claude's constitution vs Microsoft's code of conduct

Three flagship AI makers now publish behavior rulebooks:

Sources: OpenAI's Spec, Anthropic's constitution announcement of January 22, 2026, and Microsoft AI's code of conduct page, which says the draft is "not using it to train our models today" and that a revised version will guide model development in 2027. Microsoft's code does not cover OpenAI or Anthropic models that appear in Microsoft Copilot, because it applies only to MAI models.

The documents overlap a lot. All three put human oversight near the top, rank instructions from company to operator to user, and treat tool outputs as untrusted. They differ in tone. Anthropic writes mainly to Claude and explains its reasoning at length. OpenAI's Spec reads like a policy manual with test cases. Microsoft's draft insists its models should not claim feelings and rejects AI welfare and rights. These rulebooks sit alongside, not inside, each lab's catastrophic risk policy. Our Preparedness Framework vs Responsible Scaling Policy guide covers those.

Pros and cons of the new Model Spec

Pros

• Readers can now see why a rule exists, not just what it says, which makes disagreements easier to argue.
• Explicit rules on corrigibility and on behaving the same during tests give outside researchers something concrete to check.
• The memory and privacy guidance answers real complaints about assistants bringing up personal details in the wrong place.
• It stays CC0, with every past version archived, so changes are easy to track.

Cons

• It describes targets, not guarantees. OpenAI says plainly that its models do not yet fully follow it.
• More weight on judgment and inferred intent means behavior is harder to predict than under fixed rules.
• It is written and enforced by OpenAI alone, with no outside audit of how closely models comply.
• Removing most style rules leaves tone up to personalization, so two users may see quite different answers.

What it means for you

If you use ChatGPT, the practical points are simple. You can override many defaults by asking directly, such as tone, format or which side of a debate to argue. ChatGPT should tell you when it uses something from memory and should not reveal private details in drafts for other people. When ChatGPT acts as an agent, expect it to confirm scope and spending before it acts. For safer everyday habits, see how to use AI chatbots safely. If you build on the API, your developer messages outrank user messages but never the root or system levels. You should also put untrusted data in quoted blocks (OpenAI suggests YAML, JSON or XML) so the model treats it as data. For the bigger picture, our what is AI safety explainer shows where documents like this fit, and the Alignment Forum is where researchers debate them.

FAQ

What is the OpenAI Model Spec?

It is OpenAI's public document describing how its GPT models should behave in ChatGPT and the API. It sets an order of authority for instructions, firm red lines and many defaults that users and developers can change. OpenAI trains its models toward it.

What changed in the Model Spec on October 29, 2026?

OpenAI reorganized it around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. It added sections on using context and memory and on supporting human oversight. It made controlling side effects a user level principle, removed most style guidance and renamed the model "GPT."

Is the Model Spec on GitHub?

Yes. The Markdown source, a changelog and archived HTML versions since February 12, 2025 are in the openai/model_spec repository. The text is public domain under CC0 1.0, so anyone can copy or adapt it.

Does ChatGPT actually follow the Model Spec?

Partly. OpenAI says its production models "do not yet fully reflect" the Spec and that it is testing new training methods to teach the values, with results promised in future updates. Treat the Spec as a statement of intended behavior.

How is the Model Spec different from Claude's constitution?

Both are public CC0 documents used in training. OpenAI's Spec centers on a chain of command and five values with many example conversations. Anthropic's constitution, published in January 2026, ranks four priorities (broadly safe, broadly ethical, following Anthropic's guidelines, genuinely helpful) and explains its reasoning mainly to Claude itself.