OpenAI released a new version of the Model Spec (a public document defining how its models should behave, v2026.10.29) on 29 10. The new version is reorganized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, promoting human autonomy, and care for those affected; the subject in the document has changed from "the assistant" to "GPT," and most detailed stylistic guidelines have been removed.
Confidence: High (80) · Event Time 2026-10-29
Reasoning
- This is the alignment target text used when OpenAI trains ChatGPT and API models; the changes directly indicate how it wants models to make trade-offs. The official changelog lists new content including principles to help users understand, choose, and act, and adherence to relevant social norms; elevating "controlling and explaining side effects" to a user-level principle, clarifying authorization and trade-offs; and explicitly requiring support for human oversight and correction.
- A new section titled "Consider context carefully" has been added, stipulating that models must use context, memory, and personalization information in a relevant and moderate manner, while respecting privacy and user control, corresponding to ChatGPT's growing memory and personalization features.
- The specification text states that the five values are used to address two sets of conflicts: individual versus collective interests, and immediate instructions versus long-term goals; models should neither follow instructions literally only nor privately decide what users "truly want" on their behalf.
Usage
- The full text is available at model-spec.openai.com, with Markdown source files and changelogs for each version on GitHub at openai/model_spec, released under CC0 public domain, allowing direct adaptation into model behavior specifications for your own products.
Risk
- The request to the original OpenAI blog post was denied (403); this entry is based on the full official specification, the GitHub changelog, and RSS summaries, without access to the explanations and examples in the blog post.
- The specification itself states that production models do not yet fully embody the Model Spec; the public version may not include all details. Changes to the specification do not mean model behavior changes immediately.
Open · Full Model Spec (2026-10-29 version) · GitHub openai/model_spec changelog