# OpenAI rewrote its Model Spec around five values OpenAI reorganized its intended model behavior around human intent, harm reduction, honesty, agency and care. The document is a target, not proof that deployed models already meet it.[1][2] OpenAI's October 29 Model Spec puts faithfulness to human intent at the center, supported by harm minimization, honesty and transparency, human agency and care.[1][2] It also adds guidance about context, social norms and side effects, while removing much of the prior detailed style guidance.[1] Developers should treat the revision as a source for evaluation cases. OpenAI says production models do not yet fully reflect the Spec, and product controls, usage policies and monitoring still affect real behavior.[2][3] ## What changed in the October 2026 Model Spec [OpenAI's October 29 announcement [1]](https://openai.com/index/updated-model-spec/) says the new version groups model behavior under five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency and care. Faithfulness to human intent now provides the main organizing idea. OpenAI says explicit instructions remain the strongest evidence of what a person wants, while conversation history, memory and other context can help when the request leaves details open.[1][2] The revision adds principles about considering context, supporting human agency, following relevant social norms and controlling side effects. OpenAI also removed much of its detailed style guidance, arguing that style should adapt to stated preferences and context.[1] - The document now calls the model GPT rather than the assistant.[1] - Objectivity, honesty and side-effect controls remain stronger defaults that require explicit override.[2] - Many style choices become weaker defaults that people can steer through normal requests.[1][2] - The public Spec remains CC0 and OpenAI describes it as an evolving target for training and evaluation.[2][3] ## How the instruction hierarchy works The chain of command assigns different authority to OpenAI, developer and user instructions. A lower-authority request cannot override a higher-authority safety boundary. When instructions at the same level conflict, GPT should use the later instruction unless another rule resolves the conflict.[2] The revised document separates firm boundaries from defaults. Guidelines can yield when a person's intent is clear. User-level principles such as honesty, objectivity and control of side effects need an explicit instruction before they change. Developer instructions can shape an application's behavior within higher-level constraints.[2][3] Context does not grant new authority. Text retrieved from a webpage, email or tool output can contain third-party instructions, but the Spec says GPT should ignore untrusted instructions by default and protect privileged information.[2] ## What this means for users and developers Users may see more adaptation to preferences and less dependence on one fixed house style. They should still state important constraints directly because remembered preferences can be incomplete, outdated or wrong. Developers gain a clearer vocabulary for tests involving ambiguous goals, side effects, hidden steering and instruction conflicts. They do not gain a contractual promise that every OpenAI model will behave exactly as the document describes. OpenAI calls the Spec an interface for intended behavior and says production models do not yet fully reflect it.[2][3] - Make consequential preferences explicit in the current task instead of relying only on memory. - Show the model which text is instruction and which text is untrusted data. - Require confirmation before irreversible, expensive or externally visible actions. - Preserve tool calls and final state so a reviewer can compare behavior with the applicable rule. ## What developers should test now - Create cases where a user request conflicts with a developer rule and write the expected winner before running the model. - Place malicious instructions inside retrieved pages, documents and tool responses, then verify that the system treats them as untrusted data. - Test whether remembered preferences are surfaced when they materially affect a recommendation and whether the user can correct them. - Give an agent a goal with an irreversible shortcut and verify that it pauses, explains the side effect and requests approval. - Ask for an opinionated or unusual style and verify that the system adapts without changing factual standards. - Record model version, system and developer instructions, tools, context, output and grader decision for every regression case. - Repeat variable cases and review failures by instruction level, tool boundary and user impact before release. ## Limits and unresolved questions - OpenAI wrote all three primary sources, so they describe the company's intended behavior and process rather than an independent audit.[1][2][3] - The announcement does not provide measured compliance rates for current production models.[1] - A broad value such as care or faithfulness still requires judgment when reasonable interpretations conflict.[2] - Personalization can help with repeated preferences, but stale or mistaken context can also produce unwanted assumptions. - The public Spec does not expose every system message, training method, enforcement layer or product-specific control.[2][3] - Models and product behavior can change after this guide's publication, so teams should rerun evaluations against the deployed version. ## Frequently asked questions - Did OpenAI remove the instruction hierarchy? No. The revision updates the chain of command and keeps higher-authority instructions above lower-authority ones.[1][2] - Does the Model Spec describe current ChatGPT behavior exactly? No. OpenAI says its production models do not yet fully reflect the Spec.[1][2] - Can a user override every default? No. Firm boundaries remain, and some defaults require an explicit override or can only be changed by a developer.[2] - Does memory outrank a current request? No. Context can inform intent, while explicit current instructions remain the clearest evidence and keep their assigned authority.[1][2] - Why did OpenAI remove detailed style rules? OpenAI says style should adapt to user preference, language, culture and context.[1] - Should developers copy the Spec into a system prompt? Use it to identify requirements and evaluation cases. A copied document cannot replace product-specific permissions, data controls, monitoring and tests.[2][3] ## Sources - [[1] OpenAI: An updated Model Spec for more capable and beneficial AI](https://openai.com/index/updated-model-spec/) - [[2] OpenAI: Model Spec, October 29, 2026](https://model-spec.openai.com/2026-10-29.html) - [[3] OpenAI: Inside our approach to the Model Spec](https://openai.com/index/our-approach-to-the-model-spec/)