OpenAI published an updated Model Spec on 29 October 2026, and the most useful sentence in it is an admission rather than a commitment: the document describes the behaviour the company is working toward, and its models do not yet fully reflect it.

That is an unusual thing to volunteer. It is also the correct way to read the document, and the reason it belongs in a procurement file as a design statement rather than as evidence of control.

What OpenAI actually changed

The revision reorganizes the guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI says the five articulate commitments already reflected across the Spec, and that grouping them explains why the document asks models to behave as it does.

The stated rationale is capability. As models get smarter and take more consequential actions, OpenAI writes that they need to understand human intentions and values rather than only interpret the specific instructions they receive, so they can exercise judgment where instructions are silent.

The structural changes are specific. The update adds guidance on context, oversight and agency, adjusts some behavioural defaults, and retires much of the detailed style guidance. It adds a subsection on considering context carefully, defaults on supporting human agency and following relevant social norms, and expanded guidance on controlling and communicating side effects. The model is now called "GPT" rather than "the assistant".

OpenAI also concedes a limit on scope: the values are not intended to capture the full range of human values its models should respect and help people pursue.

The authority chain, stated plainly

One passage does more work than the value names. OpenAI writes that it "writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; and users direct the model's work subject to those constraints."

That is a hierarchy, and it is testable. It says a developer building on the platform inherits boundaries it did not set, and passes constraints down to users who cannot override them.

For an enterprise, the practical question follows immediately: which parts of this arrangement appear in the contract, and which appear only on a web page that can be revised without notice?

Defaults are not commitments

The second passage sets force. The Spec, OpenAI writes, turns the values into practical guidance "with firm boundaries where needed to protect people and society, and otherwise mostly takes the form of defaults that users and developers can override."

A default is a starting position. An organisation that has configured it away has no remedy in the document, and a policy that cites the values without mapping them to settings describes an intention rather than a control.

Anyone presented with the Spec as proof of governed behaviour should ask a narrow question: which parts are boundary, and which are default?

The concession worth quoting internally

OpenAI states directly that "the Model Spec describes the behavior we are working toward; our models do not yet fully reflect it." It calls the update a first step, and says it has been exploring new training methods to help models learn these values and the reasoning behind them, with results and lessons to be shared in future updates.

Read that as stated. No evaluation results, red-team outcomes or adherence figures accompany the update, and none is asserted here.

One consequence is easy to miss. If most detailed style guidance is retired, output consistency becomes a configuration choice rather than a vendor promise. Teams whose downstream process depends on a stable format now own that stability themselves.

One default that was kept, and why it matters

Not everything was retired. OpenAI retains stronger defaults, including on objectivity: users and developers can explicitly ask the model to take a particular perspective, but the model should not infer that preference merely from what it thinks they want to hear.

That is a stated position on a specific complaint, not a measured fix. Whether an assistant actually resists agreeing too readily is exactly the kind of behaviour the document says to test rather than assume.

Where judgment enters: context and memory

The new subsection on context is the part with the longest operational shadow. It explains how conversations, memories and other contextual evidence should inform a model's understanding of a person's intentions and preferences.

That is a data governance question wearing a research label. Memory is an input that shapes judgment, so retention rules, permission scope and deletion behaviour for memory stop being personalisation settings and become control settings.

One illustration shows the texture. Helping someone compare job offers, the model might draw on priorities they shared before, such as predictable hours, "while making those assumptions visible and easy to correct."

Visible assumptions are a design property, and a reasonable thing to demand. A reviewer needs to check the reasoning, not only the answer.

The same week, on the output side

Google's Gemini post describes generating, previewing and editing documents, PDFs, presentations, spreadsheets and code inside the app, with PDF preview and editing on web and mobile and other formats on web only, available to all users in the languages where the app is offered.

A separate post describes Gemini Notebook adding hundreds of textbooks from providers in the U.S. and India, including OpenStax, Open Textbook Library and S Chand Publishing, at no charge, with every answer grounded directly in the textbook, inline citations and direct links back to the original source.

Set the two kinds of output beside each other. A request to build a spreadsheet is underspecified, so the model exercises judgment, and the file carries the result without the reasoning. A grounded answer carries a link back to the page it came from.

Only the second arrives with its own defence. That is not a criticism of either feature; it is a statement about what a reviewer can reconstruct after the fact.

What none of this establishes

  • No evidence of behaviour change. No independent evaluation or adherence measurement was published with the update.
  • No contractual force described. The Spec is not presented as a service-level agreement, audit right or warranty.
  • No monitoring metrics. Nothing in the retrieved text quantifies detection, response time or incident rates.
  • No availability commitments. Google's posts describe capabilities and, for the file editor, language availability; they do not set out plan, licence or administrative constraints.

Five questions for a vendor citing a behaviour spec

  • Boundary or default? For each behaviour we rely on, is it a firm boundary or a default we could override, or lose through misconfiguration?
  • Where is the record? When a default is overridden, what captures the setting, the system prompt and the tool permissions, and who owns that record?
  • Is it incorporated? Does the agreement reference the Spec, or merely link to a page that changes on the vendor's schedule?
  • What feeds judgment? Which contextual inputs shape output, and what are their retention and permission rules?
  • What travels with the artifact? When output becomes a file someone else edits and signs, what provenance accompanies it?

Why this matters more than a model launch

Three observations follow, and they are analysis rather than reporting.

First, the control surface has moved from policy to configuration. If most of a behaviour specification is overridable, the enforceable control is the set of choices an organisation actually made, plus the record of who made them.

Second, the admission is more useful than the aspiration. A document that says the models do not yet fully reflect it tells a buyer what to test and what to monitor, which is more than a claim of compliance would offer.

Third, judgment in underspecified situations is now the product and the risk simultaneously. The same capability that lets a model infer what a person meant is the capability that produces a confident artifact nobody can trace.

Evidence notes and limits

Evidence was reviewed through 15:00 UTC on 30 October 2026. An earlier automated retrieval of the OpenAI Model Spec post returned HTTP 403; the page was subsequently retrieved with substantive text, and the five values, the authority quotation, the defaults statement, the retained objectivity default, the "do not yet fully reflect it" concession, the structural changes, the job-offer illustration and the reference to training work with results promised in future updates all come from that text.

The two Google posts are described at the level of their published descriptions. Their capability and availability language is Google's own, and no performance, adoption or outcome figure is attributed to them.

No customer results, benchmarks, financial results or stock movements are asserted. The five procurement questions and the closing analysis are the author's, clearly separated from vendor statements.