OpenAI Model Spec Rewrites Around Five Values for GPT

What changed

OpenAI published a full Model Spec update on 29 October 2026. The document is the public rulebook for how GPT models should behave in products and APIs. This release reorganizes the Spec around five values. It also renames the governed product from “the assistant” to GPT.

OpenAI posted a blog post, release notes, and a dated HTML snapshot on model-spec.openai.com. The GitHub source repo tagged release 2026-10-29 on the same day. The rewrite removes most detailed style guidance. It adds new principles on side effects, memory, personalization, and human oversight.

The Spec still says misalignment happens when a model follows goals that fight higher level instructions. The chain of command section explains which messages can override which rules.

Why it matters

Safety and product teams need a stable reference when they design evals, logging, and escalation paths. Before this release, teams often inferred rules from scattered blog posts. Now they can map features to named values and test for gaps.

The Spec also states plainly that production models do not yet fully match the document. That line matters for regulated buyers. You should not claim full alignment in contracts unless your own evals prove it.

Who is affected

Leads for applied AI, trust and safety, and legal review at enterprises using OpenAI tools. Independent eval vendors may extend Model Spec suites already published on GitHub. Competitors may copy the five value frame in their own policy docs, which changes how buyers compare vendors.

Support and developer relations teams should update internal playbooks. Public docs should use the word GPT where the Spec does, so customers are not confused by old “assistant” language.

What to do next

Download the 29 October HTML snapshot and compare it to your current policy checklist. Add Spec sections to regression evals if you run automated safety tests. Brief customer facing teams on renamed terms and on new guidance about memory and side effects.

If you log tool calls, review the user level principle on communicating side effects. Make sure your product surfaces risky actions before they run.

What to watch

OpenAI said the Spec will keep changing based on feedback. Watch for the next dated HTML release. Watch the Model Spec evals repository for new prompt packs that track the five values.

Also watch whether regulators cite the Spec in AI governance guidance outside the United States.

Photo: RDNE Stock project / Pexels

The five values are faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. The overview explains how GPT should read intent when instructions conflict or when context is incomplete.

Developers who fine tune or route models should read how the Spec treats developer messages versus end user messages. That order decides which instruction wins in a conflict.

Photo: cottonbro studio / Pexels

OpenAI hosts evaluation tooling that scores model answers against Spec sections such as chain of command and staying in bounds. Teams can reuse that harness instead of writing prompts from scratch.

The changelog for v2026.10.29 lists each major section move. It links the blog post and help center release notes for readers who want a short summary before reading the full markdown source.

Policy leads should treat the Spec like an API contract for behavior. When your app sends system prompts, tool definitions, and memory snippets, you are stacking instructions with different authority levels. The updated chain of command section is the map for that stack.

If you run red teams, save the October HTML version as a baseline. Future edits may be small, but dated snapshots make audits easier. Share the five value names with UX writers so error messages and refusals use the same words customers see in OpenAI docs.

The Spec is public domain under CC0. That choice invites forks and comments from outside OpenAI. Your team can diff the markdown in Git without paywalls. Use that openness to build internal checklists that cite section titles verbatim.

Models will not match the Spec perfectly on day one. OpenAI says so in the overview. Plan eval cycles rather than one time sign off. Re test after each model upgrade you enable in production.

Sources

Primary — OpenAI, Updated Model Spec blog post (29 October 2026). Announces the five value structure and GPT naming.
Primary — OpenAI, Model Spec changelog on GitHub (29 October 2026). Lists concrete section changes for release v2026.10.29.
Secondary — OpenAI, Model Spec source repository (29 October 2026). Hosts markdown source and archived HTML versions.