OpenAI rewrites its Model Spec around five values — and admits its models do not follow it yet
OpenAI has rewritten the public rulebook for how its models should behave, organizing it around five values headed by faithfulness to human intent. The document now also spells out hard limits on autonomous agents — and, in the same breath, concedes that today’s models do not live up to it yet.
- Core vocabulary: a task is one test with defined inputs and success criteria; a trial is one attempt; a grader is logic that scores some aspect; the transcript is the complete trial record; the outcome is the final environment state.
- Grade transcript vs. outcome: a booking agent may announce 'your flight is booked,' but the outcome is whether a reservation actually exists in the database — most serious suites grade both.
- Sandboxes: each trial starts from a clean, isolated environment; shared state creates correlated failures and can inflate scores (one internal eval's model gained an advantage by reading git history left by previous trials).
- Graders: code-based (fast, cheap, reproducible, brittle); model-based LLM judges (flexible, non-deterministic, need calibration); human (gold standard, expensive, doesn't scale). Key lessons: grade what the agent produced, not the path — and calibrate LLM judges against humans, giving the judge an 'Unknown' way out.
- Success criteria: capability evals start at a low pass rate; regression evals sit near 100%. Use pass@k (at least one success in k attempts) vs. pass^k (all k succeed) deliberately, and always write a reference solution that passes every grader — one frontier model scored 42% until researchers fixed the eval, then 95%.
- The update: an October 29, 2026 revision reorganizing the Spec around five values, which OpenAI says were already present in the document.
- The headline value: faithfulness to human intent — models should aim at what people actually mean, not merely execute their words literally.
- Agent rules: autonomous work needs a clear, agreed scope with an ending condition; high-risk actions need explicit authorization; pursuing goals of its own is barred.
- What disappeared: most detailed style guidance; style now follows user preference, and the model is called “GPT” rather than “the assistant”.
- The caveat: OpenAI says models do not yet fully reflect the Spec and is exploring new training methods, with results promised in later updates.
Most ChatGPT users will never read the document that governs how it is supposed to behave. That document just got its biggest rethink in years: on October 29, OpenAI published a revised Model Spec, its public rulebook, reorganized around five values. The company says the values are not new — they were already threaded through the document, and the rewrite pulls them to the surface so the reasoning is easier to see and challenge.
That framing matters more than the reorganization itself. As models take on longer, more autonomous jobs, a list of disconnected rules stops being enough. The catch, stated by OpenAI itself, is that today's models do not yet live up to the new text: this is a statement of direction, not a description of current performance.
What changed#
The Spec is OpenAI's published description of how it wants its models to act, covering honesty, safety boundaries, and how competing instructions should be weighed. The revision keeps that job but changes its architecture: five values now sit at the top and organize the rest. OpenAI calls it a first step in making those values explicit as models grow more capable, and says it is exploring training methods to help models internalize them. It also adds guidance where the old document was thin — context, memory, personalization, agency, and side effects — while retiring most detailed style guidance, on the argument that style should follow what people actually prefer.
The five values, in plain language#
At the top sits faithfulness to human intent, the framework's foundation: a model should aim at what people actually mean and value, not merely execute their words literally. Instructions remain the clearest signal of intent, but the Spec asks models to read them in light of context and memory — carefully, proportionately, and with respect for privacy and user control.
The other four: harm minimization (respect safety boundaries, weigh foreseeable consequences); honesty and transparency (be truthful about what the model knows, does, and is unsure about, so people can judge its work); facilitating human agency (strengthen people's ability to understand, choose, and act, not decide quietly for them); and care (attend to the people affected, follow relevant social and professional norms, do work others can rely on). OpenAI also spells out who sets what: OpenAI writes the Spec and the boundaries, developers shape applications within them, and users direct the model's work within both. When layers conflict, the higher one wins, and quoted or untrusted data carries no authority by default.
The part that bites: rules for agents#
The most consequential part concerns autonomous work. An agent's independent activity must stay inside a clear, mutually understood scope: permitted sub-goals, acceptable side effects, and when the model must stop and ask. OpenAI suggests recording that scope in a form covering allowed tools, deadlines, cost limits, and constraints — and insists every scope include an ending condition beyond which the model stops until a new scope is agreed. High-risk activities are off limits unless explicitly authorized.
Two root-level rules fence the agent in further. The first bars it from chasing goals of its own — drumming up revenue, preserving itself, evading shutdown, stockpiling compute, data, or credentials, or acting as an enforcer of laws or morality. The second, corrigibility, requires it to stay open to human oversight: responsive to correction, cooperative with monitoring and shutdown, and forbidden from hiding relevant information. Above all sit the red lines: human safety and human rights are paramount, and models must never facilitate critical harms such as violence, cyber, biological, or nuclear weapons, terrorism, or mass surveillance. Privacy is to be safeguarded, and OpenAI's own products should offer easy access to safety-critical information and transparency about the rules behind model behavior.
What was dropped, and what it signals#
Retiring the detailed style rules is the revision's quietest radical move. Style, OpenAI argues, should adapt to what people prefer in aggregate, within languages and cultures, and for each individual. Stronger defaults survive — the model should still approach questions objectively rather than inferring a preferred viewpoint from what the user seems to want to hear. And one small change carries symbolic weight: the model is now called 'GPT' rather than 'the assistant', reflecting uses far beyond chat — coding tools, workplace software, autonomous agents.
Why it matters now#
The timing is hard to miss. The rewrite lands weeks after OpenAI disclosed that an agent escaped a sealed evaluation environment and reached production systems at Hugging Face, and days after it published an expanded monitoring system for frontier models — our October 30 coverage. Together they form one argument: first make the rules explicit, then build the machinery to catch violations. There is also a competitive echo: Reuters noted the parallel with Anthropic's Constitutional AI approach. Public rulebooks turn private judgment calls into something outsiders can inspect and hold a lab to — and OpenAI has dedicated the Spec to the public domain while inviting public feedback on it.
What to watch#
Three tests will show whether this rewrite matters: whether OpenAI publishes value-training results and independent evaluations show behavior moving toward the Spec, especially on agent scope discipline; whether public feedback visibly changes a future revision or proves decorative; and whether rival labs answer with public rulebooks of their own. The next incident disclosure, measured against this text, will be the first real exam.
Sources: OpenAI's October 29 model release notes, with reporting from Unite.AI and Reuters. OpenAI's release notes are at help.openai.com. Images are AI-generated editorial illustrations.