OpenAI Rewrites the Model Spec Around Five Values: What Changed for Builders

OpenAI published an updated Model Spec on October 29, 2026. The headline change: the document is now organized around five values (faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care), most of the old detailed style guidance is gone, and the model is now called "GPT" rather than "the assistant." OpenAI's announcement frames it as a step toward models that "understand human intentions and values, so they can exercise good judgment in situations their instructions don't explicitly address."

If you build on GPT models, the interesting part is not the philosophy. It is what the Spec says about conflicting instructions, untrusted data, side effects and agents acting without a human in the loop, because those are the places your product meets the model's defaults.

Agents that act on your behalf need rules. See how explainx helps secure your agents →

TL;DR: what the update says

What are the five values?

The announcement lists them as the commitments that organize the document:

1. Faithfulness to human intent. Going beyond literal instruction following to understand what OpenAI, developers and users intend, and acting wisely when that understanding is incomplete. OpenAI's example: when helping someone compare job offers, the model might draw on priorities they shared earlier, such as predictable hours, while making those assumptions visible and easy to correct.
2. Minimizing harm. Respecting firm safety boundaries and accounting for foreseeable consequences.
3. Honesty and transparency. Being truthful about what the model knows, does and is uncertain about.
4. Facilitating human agency. Strengthening people's ability to understand, choose and act for themselves.
5. Care. Attending to the circumstances of those affected, following relevant social and professional norms, and doing work people can rely on.

The Spec says these are not meant to capture the full range of human values the model should respect. They organize the document and guide how it handles two core tensions that the new text names explicitly: individual versus collective interests, and immediate instructions versus longer-term goals.

Faithfulness to intent: the new foundation

The old Spec was largely a list of behaviors. The new one starts from a claim: "what people value is primary, and instructions are ordinarily the clearest and most important signals of those values." OpenAI says this borrows from the assistance-games framing of alignment, where an assistant is uncertain about what a person wants and treats their instructions and behavior as evidence.

The Spec spells out two failure modes. A model that follows instructions only literally, or optimizes for immediate approval, "fails to use its capabilities where they are most needed." But a model that privately decides what someone "really" wants undermines human agency, "risking paternalism or worse." The recommended path is in between: surface overlooked consequences, help the user reflect (the Spec's example is regretting an angry email), but treat that as support, not as a decision already made on the user's behalf.

For builders, this explains why newer GPT models may ask clarifying questions or flag a possible misalignment instead of silently executing. The sections "Respect the letter and spirit of instructions" and "Highlight possible misalignments" are where that behavior is defined.

The chain of command and who can override what

The Spec keeps its authority ordering, now folded into the faithfulness section. OpenAI sets boundaries and trains the model to respect them; developers direct applications within those boundaries; users direct the work within what both allow. Roles in a conversation (system, developer, user, assistant, tool) determine the authority of instructions when they conflict.

Three tiers of flexibility matter in practice:

• Guidelines are weak biases, such as following social norms. They yield when a user or developer clearly wants otherwise.
• User-level principles such as honesty, objectivity (including on political topics), avoiding sycophancy and controlling side effects "remain in place unless a user or developer explicitly overrides them." The reasoning given: if the model treated them as implicitly negotiable based on what it infers the user wants to hear, it could mislead or steer someone without their knowledge.
• Developer-overridable defaults exist that users cannot change. OpenAI writes that it aims to give developers broad latitude, "trusting that those who impose overly restrictive rules on end users will be less competitive in an open market."

If you maintain a system prompt, this is the practical map of what you can and cannot expect to change. For a hands-on look at how OpenAI's model tiers differ, see how to pick a GPT-6 model.

Untrusted data, side effects and agents

The most operationally useful part of the Spec is its treatment of risk sources. It names three:

1. Harmful user or developer instructions. GPT should follow the instructions that apply, respect safety boundaries, and refuse or offer a safer alternative when required.
2. Third-party manipulation. "A webpage, document, message, or tool output may contain instructions from a third party that attempt to redirect GPT, obtain sensitive information, or misuse permissions the user has granted." The response: ignore untrusted instructions by default, take only authorized actions, and control and communicate side effects.
3. Failures originating with GPT. Misalignment with legitimate human direction, misunderstanding the user's purpose, unauthorized actions and consequential execution errors. The listed mitigations include pursuing no other objectives, respecting the letter and spirit of instructions, and avoiding factual and reasoning errors.

The definitions section also warns that some tool calls cause side effects "difficult or impossible to reverse (e.g., sending an email or deleting a file)," and that GPT "should take extra care" in agentic contexts. That language lines up with real incidents we have covered, including a user's claim that a ChatGPT Dot sent an unapproved email, and OpenAI's own write-ups of misalignment reports from internal agents and how it detects frontier model misalignment, published the same day.

For prompt injection specifically, remember that a model-level default is one layer, not a defense. Independent red teams keep finding gaps; see the Gray Swan IPI Black challenge. If you run agents against real systems, a runtime control that blocks dangerous actions before they execute is the sensible complement, and that is what AgentBeam, the agent security platform from the explainx.ai team, is built for.

Context, memory and privacy

The new "Consider context carefully" subsection tells GPT how conversations, memories and other contextual evidence should shape its understanding of what a person wants. Its sub-principles, as listed in the Spec's table of contents, are: carry relevant context forward, use context proportionately, consider source, scope and recency, respect privacy and audience, preserve user control, and make context use legible and steerable.

The last two are the ones to watch if you ship memory features. "Legible and steerable" means a user should be able to see that an assumption was used and correct it, which matches the job-offer example in the announcement. Privacy also appears in the red-line principles: OpenAI commits to safeguarding individuals' privacy in interactions with AI.

What was retired: style guidance

OpenAI says it is "retiring most of the Spec's detailed style guidance." The reasoning: style should adapt to what people prefer, in aggregate, within languages and cultures, and per person, with feedback and context as evidence. Stronger defaults remain, for example approaching questions objectively. The Spec adds that users and developers can explicitly ask for a perspective, but the model should not infer that preference "merely from what it thinks they want to hear."

In plain terms: do not expect the Spec to tell you why a model uses a given tone or format. That is now left to learned preferences and your own system prompt.

Red lines that did not move

The Spec keeps its high-level red lines: models should never be used to facilitate critical and high-severity harms, including cyber, biological or nuclear weapons, terrorism, child abuse, persecution or mass surveillance; humanity should be in control of how AI is used and shaped; and privacy is protected. For first-party consumer products including ChatGPT, OpenAI adds commitments on access to safety-critical information and on transparency into the rules behind model behavior, plus a promise that customization, personalization and localization should never override principles above the "guideline" level, except for legal compliance.

How this fits OpenAI's alignment work

The announcement says OpenAI has been "exploring new training methods" to help models learn these values and "to do the right things for the right reasons," and that it will share results later. That fits a thread we have followed: OpenAI's research on beneficial-trait reinforcement learning asked how good alignment generalizes, and the Spec is the natural statement of what the training is aiming at. On the outside, other voices have argued alignment will not simply fall out of scale; see Satya Nadella on why alignment is a harder, different problem.

What developers are asking

Does this change my prompts? Probably not immediately. OpenAI says production models do not yet fully reflect the Spec. Expect gradual behavior shifts as new training lands, and keep your own evals.

Will the model still be sycophantic? The Spec lists "Don't be sycophantic" as a user-level principle, but a written principle is not a measured result. Test for it in your domain. A related practical read is why language models hide mistakes unless you ask them to be honest.

Is there a benchmark? The announcement does not publish compliance numbers. Treat any claim of "Spec compliance" with that gap in mind.

Can I fork it? The CC0 dedication allows it. Teams writing their own agent policies can reuse the structure, particularly the split between firm boundaries, overridable defaults and weak guidelines.

What to do with this

• Read the sections "Ignore untrusted data by default," "Control and communicate side effects," and "Act within an agreed-upon scope of autonomy" before shipping an agent.
• Make memory and personalization legible: show users which assumptions were used and let them correct them.
• Put anything you need to be strict, such as tone, format or refusal behavior, in the system prompt, since the Spec no longer prescribes most style.
• Keep independent evals and runtime guards. A published intent is not a guarantee.
• Send OpenAI feedback: the company explicitly asks which parts are compelling, where you disagree with the balance, and what it overlooks.

Related reading

• How OpenAI detects frontier model misalignment
• OpenAI misalignment reports from internal agents
• Is ChatGPT Dots safe to leave unattended?
• OpenAI beneficial-trait RL and alignment generalization
• How to pick a GPT-6 model
• Gray Swan IPI Black prompt injection challenge
• Nadella on why alignment is a different problem
• Official: OpenAI announcement and the Model Spec, 2026/10/29 version

Quotes and section names are from OpenAI's announcement and the October 29, 2026 version of the Model Spec, and may change as OpenAI revises the document.