OpenAI updates the Model Spec for the first time since December

OpenAI published a new version of the Model Spec dated August 18, 2026, its first revision since December 18, 2025. The changelog lists added clarity on appropriate relational interactions for teens, clearer handling of requests built on false or unsupported premises, removal of outdated guidance on how pre-reasoning models should think through a request, and a new section titled 'Be clear about capabilities and limits'.

Why it matters

The Model Spec is the document OpenAI's models are trained and evaluated against, so it is the closest thing to a written contract for how ChatGPT and the API behave. An eight-month gap between revisions means these changes describe behaviour that is already shipping rather than a proposal open for comment.

Who should care

Teams building on the OpenAI API, and anyone writing safety or escalation policy around an assistant

What you can do

Read the new version at model-spec.openai.com before assuming older prompt workarounds still hold, particularly around false-premise handling and how the model is expected to describe its own limits.