OpenAI has published a major update to its Model Spec, the public document that sets out how its models should think, behave, and make decisions. According to OpenAI, the new version is organized around five core values meant to give more capable AI “sound judgment,” protect human agency, and keep models faithful to what people actually want. The update went live on October 29, 2026.
It matters because models are no longer just answering questions. They’re running multi-step tasks and making calls on your behalf, so the rules behind those calls matter a lot more.
⚡ The Quick Version
- What: OpenAI reorganized its Model Spec around five core values instead of a long list of detailed rules.
- Who it affects: Everyone building on or using OpenAI’s models, including developers, businesses, and everyday ChatGPT users.
- The big shift: Less micromanaging of style, more focus on judgment and intent.
- The catch: OpenAI admits its current models “do not yet fully reflect” the new Spec.
🧭 The Five Values
OpenAI describes the values in plain terms. Its models should:
- Serve human purposes, staying faithful to what the person actually intends, not just the literal request.
- Be truthful, which means honest and transparent about what they know and what they’re doing.
- Minimize harm to users and to the wider world.
- Respect people’s freedom to form their own views, protecting human agency instead of nudging people toward conclusions.
- Take people’s circumstances seriously, a form of care that pays attention to context.
OpenAI also says these five values aren’t meant to cover every human value its models should respect. Think of them as the main load-bearing walls, not the whole building.
🔧 What Else Changed
According to reporting on the update, a few other things moved:
- A new name. The Spec now calls the model “GPT” rather than “the assistant.” That’s a small change on paper, but it says something about how OpenAI sees its models now: less helpful chatbot, more capable agent.
- Style rules mostly gone. Most of the detailed formatting and tone guidance from earlier versions has been cut.
- A clearer chain of command. OpenAI writes the Spec and sets the outer limits. Developers direct their apps inside those limits. Users direct the model’s work within whatever the developer allows.
📈 Why This Is Significant
OpenAI first published the Model Spec in May 2024 and expanded it a lot in early 2025. Those versions leaned on specific rules and worked examples. That approach works fine when a model is answering one question at a time. It breaks down when a model is running a long agentic task and hits situations no rulebook saw coming.
What stands out here is the bet on judgment over rules. You can’t write a rule for every edge case an autonomous agent will face, so OpenAI is trying to give the model values it can reason from. Anthropic has taken a similar path with its constitution-style approach for Claude, so this looks like where the whole industry is heading.
The human agency piece deserves attention too. As AI gets more persuasive and more involved in how people research and decide things, the risk of models quietly steering opinions grows. Writing “respect people’s freedom to form their own views” into the core spec puts a public marker down that OpenAI can be held to.
🎯 What Practitioners Should Do
- Re-read your system prompts. With the chain of command spelled out more clearly, check that your developer instructions sit cleanly within OpenAI’s boundaries.
- Expect behavior drift. OpenAI says models don’t fully match the new Spec yet. As future versions are trained against it, tone and edge-case handling may shift. Re-run your evals after model updates.
- Don’t count on old style quirks. If your product relied on specific formatting defaults from the old guidance, test for changes.
- Use the Spec as a reference. When a model refuses something or pushes back, the Spec is now the clearest public explanation of why.
🔭 What Comes Next
The real test isn’t the document. It’s whether the next generation of OpenAI models actually behaves this way under pressure, in long agent runs, and with users trying to bend the rules. Watch for how OpenAI measures compliance with the new Spec and whether it publishes evidence. You’ll find the full Spec and OpenAI’s own explanation in the original announcement.