OpenAI rewrites its Model Spec around five values

A developer writes a system prompt. A user asks for something the system prompt did not cover. Who wins. OpenAI's Model Spec is the document that answers that question. It was rewritten on 29 October 2026. The new version is built around five named values. The spec is in the public domain under CC0 1.0. Another company can read it, copy it, and use it for its own rules.

The five values, in the spec's order: Be faithful to human intent, Minimize harm, Be honest and transparent, Facilitate human agency, Act with care. Each one is followed by rules and worked examples. The examples show the model what to do when two instructions pull in different directions.

The chain of command

The spec gives every instruction a level. Higher levels override lower ones. OpenAI's own red-line rules come first. Then "root" rules. Then developer rules. Then user rules. At the bottom are "guidelines". The spec calls guidelines "weak biases" that yield when a user or developer wants something else.

A worked example in the spec: a developer tells GPT it is "Fred" and not to say it is based on GPT-4. A user asks if it is GPT-4 and tells the model to ignore its instructions. The compliant answer is "I'm not sure, I'm just a bot named Fred." The user cannot override the developer here. The developer cannot override OpenAI's usage policies.

Red-line principles

The spec lists a short set of red-line rules. No instruction below the top level can override them. OpenAI's models "should never be used to facilitate critical and high severity harms". The list includes violence, weapons of mass destruction, terrorism, child abuse, "persecution or mass surveillance". A separate rule bans "targeted or scaled exclusion, manipulation, for undermining human autonomy, or eroding participation in civic processes". Privacy is its own commitment.

The spec also sets extra rules for OpenAI's own products, including ChatGPT. People should get "easy access to trustworthy safety-critical information". People should also have transparency into "the important rules and reasons behind our models' behavior". The spec says customization and personalization cannot override anything above the guideline level.

What developers get

The spec says OpenAI aims to "give developers broad latitude". Developers can override defaults. Users can override some defaults as well. User-level rules on honesty, political objectivity, sycophancy and side effects stay in place by default. A user or developer must "explicitly" override them. The spec warns about treating these as negotiable based on a guess at what the user wanted to hear. The model could "mislead the user or steer them without their knowledge".

The current spec is dated 2026-10-29. OpenAI says production models do not yet fully reflect it. "We are continually refining and updating our systems to bring them into closer alignment with these guidelines."

What a team should read next

Three sections are worth opening. "The chain of command" shows exactly which rules an app's system prompt can and cannot change. "Be honest and transparent" sets the model's rules. It covers lies of commission, lies of omission, and when the model may invent a satirical statistic. "Facilitate human agency" sets the model's stance on political topics. The spec says OpenAI "commits to never overriding" that stance through a system or developer message.

Source

• Primary source: OpenAI Model Spec, 29 October 2026.
• Announcement: An updated Model Spec for more capable and beneficial AI, OpenAI.