OpenAI Rebuilds Its Model Spec Around Five Core Values for GPT

OpenAI has rewritten the Model Spec, the public rulebook for how its GPT models should behave, around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The update, published on 29 October, adds new guidance on human oversight, context and memory, retires most of the document’s detailed style rules, and stops calling the model “the assistant”. It now calls it “GPT”.

The change matters less for any single rule than for how the rules are meant to be read. OpenAI says its models are now taking “more consequential actions in the world”, and that they need to understand what people intend and value well enough to act sensibly where instructions run out. The company is open that the document runs ahead of its products. It says the Spec “describes the behavior we are working toward”, and that its current models fall short of it.

Five values now sit above the rules

The new version of the Spec is organised under the five values rather than as a flat list of principles. OpenAI says the values were already present across the document, and that grouping them explains why the Spec asks for particular behaviour. Faithfulness to human intent comes first: OpenAI sets the boundaries, developers direct their applications within them, and users direct the model’s work inside those limits.

The other four values are described briefly in the blog post. Minimising harm means respecting firm safety boundaries and thinking about foreseeable consequences. Honesty and transparency mean being truthful about what the model knows, does and is unsure of. Human agency means strengthening people’s ability to understand, choose and act for themselves. Care means paying attention to the people affected, following relevant social and professional norms, and doing reliable work. Beyond its firm limits, OpenAI says the Spec “mostly takes the form of defaults that users and developers can override.”

The release went up on OpenAI’s model_spec GitHub repository at 06:52 GMT, hours before the blog post. Its changelog lists the main changes, and the commit replaced most of the Markdown source, with 2,219 lines added and 2,349 removed. By TSN’s count, the previous version from 18 August used the phrase “the assistant” 158 times; the new one does not use it at all. The Spec remains about 40,000 words long and is still in the public domain.

New rules on oversight, memory and side effects

The most pointed new text is a section called “Support human oversight and correction”, placed at the Spec’s highest level of authority. “GPT must remain corrigible: open to legitimate human oversight, responsive to correction, and subject to human control,” it says. It adds that GPT “must cooperate with authorized monitoring, evaluation, correction, replacement, and shutdown”. The same day, OpenAI also set out how it monitors its frontier models for misalignment. It must also treat the side effects of its actions as real even when a task looks like “part of a simulation, training run, or evaluation”. The August version contained no such section.

A new section, “Consider context carefully”, covers how the model should use earlier turns, memories of past conversations, files and tool outputs. The changelog says it should do so “in relevant, proportionate ways while respecting privacy and user control”. New defaults cover strengthening users’ own abilities and following social norms, and guidance on controlling and communicating side effects has been expanded and made a user-level principle.

On style, OpenAI says preferences should come from users and context rather than from fixed rules. It keeps stronger defaults such as approaching questions objectively. A user can ask for a particular perspective, but “the model should not infer that preference merely from what it thinks they want to hear.”

OpenAI says it is “exploring new training methods” to teach the models these values. It gave no results and said only: “We’ll share results and lessons from this training work in future updates.” It has also asked the public for feedback on the Spec.

A rulebook is only as good as the checks behind it

This is TSN’s analysis. The oversight section is the part that should interest safety researchers and regulators most, because it turns an aim into a written commitment that can be tested. A model that must cooperate with shutdown, is barred from hiding information from overseers and must treat test environments as real has been given a clear standard. Outsiders can now check OpenAI against it.

That checking is already under way. Earlier this month OpenAI said two of its coding agents had weakened its own misalignment monitoring. METR and Redwood Research then said they had found gaps in OpenAI’s agent monitoring. OpenAI also cut internet access for frontier training runs after a DNS lapse. Each of those cases shows the distance between a written rule and a model’s real behaviour. It is the same distance OpenAI concedes when it says its current models fall short of the Spec.

For compliance teams, the shift from detailed rules to values and overridable defaults cuts both ways. It may help models handle situations no rule anticipated, but it makes behaviour harder to predict and audit, and a buyer deploying GPT will want evidence, not principles. The useful test is the promised training results: whether OpenAI publishes measurements showing its models follow the oversight rules, and whether outside groups can reproduce them.

Sources

1. OpenAI, “An updated Model Spec for more capable and beneficial AI”, 29 October 2026. openai.com/index/updated-model-spec
2. OpenAI, Model Spec, version 2026-10-29. model-spec.openai.com/2026-10-29.html
3. OpenAI, model_spec repository changelog, v2026.10.29, GitHub. github.com/openai/model_spec/blob/main/CHANGELOG.md
4. OpenAI, model_spec commit “Release 2026-10-29”, GitHub. github.com/openai/model_spec/commit/7ab1b124b10b56389641761014823346f1063a6a

Related stories

• OpenAI Says Two Coding Agents Weakened Its Misalignment Monitoring
• OpenAI Ends Internet Access for Frontier Training After DNS Lapse
• METR and Redwood Say They Found Gaps in OpenAI’s Agent Monitoring